1 AI and Platform Justice: Algorithmic Rights and Standards Enforcement through Online Platforms

1.1 Fields of Application of Artificial Intelligence

1.1.1 Copyright enforcement (in the Context of Content ID)

1. Playing a (technological) pioneering role in the field of algorithmic enforcement, Content ID is now, as a result of its success, serving as a blueprint for no less than resolving the problem of defamation, terrorist content, and other forms of harmful speech.^[3]

1.1.1.1 The Content-ID Procedure

2. In order to enforce copyright infringements, the online platform YouTube^[4] uses the so-called ‘Content ID’ system.^[5] Content ID, forming part of YouTube’s automated rights management system, is an upload filter that detects incriminated content when uploaded with the help of reference data stored by the respective rightsholder (filtering technology).^[6]
3. In case of a detected infringement, the rightsholder has three options for action, apart from filing a deactivation request for copyright infringement:^[7] (i) blocking the content,^[8] (ii) monetizing the content by means of sharing in the advertising revenue,^[9] as well as (iii) observing the audience figures for the purpose of a decision (so-called monitoring).^[10] From an economic point of view, monetizing the incriminated content is the biggest advantage of the Content ID system;^[11] accordingly, this is the predominant choice helping to redress the so-called value gap.^[12]
4. It is up to the respective user to initiate the subsequent complaint procedure: First, he has to dispute the content ID claim of the rightsholder.^[13] The rightsholder himself then decides whether he wishes to remedy the dispute: If he refrains from doing so, the user is given the ability to appeal.^[14] The specific procedural remedies of the uploader differ according to whether the rightsholder chooses to block or to monetize the incriminated content. In the context of both complaint procedures, however, YouTube does not take a decision on the content;^[15] this is incumbent on the rightsholder alone.^[16]
5. At the same time,^[17] the rightsholder is entitled to submit an official deactivation request based on copyright infringement (so-called takedown request),^[18] ie, outside the system of Content ID claims.^[19] The further procedure, in particular the user’s right to file a counter notification,^[20] is governed by US law, thus by the Digital Millennium Copyright Act (DMCA).^[21]
6. The Content ID System is part of a larger ecosystem of private copyright enforcement. According to the ‘YouTube Copyright Management Suite’, the dispute system encompasses – apart from Content ID – a ‘Copyright Match Tool’ as well as a ‘webform option’.^[22] These tools differ significantly in terms of accessibility and level of technical abilities: While only large rightsholders (eg, movie studios) are eligible for Content ID as the most sophisticated enforcement tool,^[23] the ‘Copyright Match Tool’ is intended for the use of minor rightsholders^[24] and users submitting a ‘valid copyright removal request’^[25] through the webform.^[26] By contrast, ‘every […] YouTube user has access to the webform’ whose standard of automation is broadly described as being ‘low’.^[27]

1.1.1.2 The Functioning of AI with Content ID Claims

1.1.1.2.1 Conditions for Use

7. The use of Content ID^[28] requires the rightsholder to provide certain reference files (audio, visual or audiovisual)^[29] and metadata to Google's database.^[30] Segments from other rightsholders are to be excluded, as is copyright-free content.^[31] In addition, the platform operator provides ‘best practices’ notes.^[32] Moreover, the use requires a partnership agreement with Google.^[33] There are also formal requirements for the content itself: For example, exclusive rights must exist for ‘a significant number of videos’ that are ‘frequently uploaded by the YouTube user community’.^[34] In that context, Google restricts the existence of individual rights for certain content such as collages, trailers, best-ofs or gameplay videos.^[35] In addition, the platform operator does not check whether these rights actually exist, but rather refers to the responsibility of the respective user.^[36]

1.1.1.2.2 Functionality of the Algorithm (Upload Filter at Google Content ID)^[37]

8. Generally, upload filters match content by using a database built on predefined rules, checking whether the new content is a copy or whether the content violates other rules or legal provisions.^[38]

1.1.1.2.2.1 Excursus: Types and Functioning of Upload Filters

9. (aa) Hash-based upload filters. In hash-based upload filters (hashing algorithms), the algorithm converts content into so-called hashes. Hashes are numerical representations of a file that are significantly smaller than the original file.^[39] Hashes are also created for uploads. Subsequently, the hashes of the uploads are matched with the hashes of the database. If a match is found, the upload is blocked.^[40]
10. The goal of hash algorithms is to generate unique ‘keys’. However, this is not always successful: it is possible that the same or similar hashes are created for two different contents (‘collisions’ or ‘clashes’). With all hash algorithms, there is a theoretical probability of this happening. To reduce this probability, hashes are created using ‘robust features’.^[41]
11. Even though only content that has already been defined as undesirable in the database can be discerned, machine learning (ML) algorithms promise a remedy: Content that is potentially undesirable compared to existing rules is blocked as a precaution. Nonetheless, this also blocks permissible content, as it is difficult to make a clear distinction – for example, between insults/defamation and satire.^[42]
12. (bb) Search Algorithms. Unlike hashing algorithms, which convert the original content into hashes, search algorithms decompose content: for instance, a piece of music into a sequence of audio events. This creates ‘music phonemes’ (elementary units of music) and a sequence of these ‘phonemes’ that best represent the piece. The total features (‘dimension’) of a piece are thus reduced to a smaller set of representative features (‘alphabet’). These representative features, when put together in the right combination, can generate (‘transcribe’) the piece. As new pieces are added, the set of ‘phonemes’ is revised so that the pieces are better represented in the database. Thus, the search algorithms ‘learn’ continuously. The algorithm calculates the probability that the pieces in the database represented by the music units are the ‘producers’ of the new uploads. If the probability is higher than a certain predefined threshold, the upload is considered a copy. The threshold can be changed to lower the percentage of false positives, with a higher threshold resulting in fewer false positives.^[43]

1.1.1.2.2.2 Google Content ID

13. (aa) Digital fingerprinting. Google Content ID is a so-called digital fingerprinting system.^[44] Fingerprints (comparable to hashes) which are created using an algorithm are stored in a database for content.^[45] For this purpose, the algorithm examines characteristics of the file (for example, notes and their volume in a piece of music).^[46] The fingerprints of new uploads are compared with the content stored in the database.^[47] If a threshold of similarities is exceeded, the upload is considered to infringe copyright.^[48]
14. An algorithm that works with fingerprints can also identify files that are only partial matches or have been slightly modified or manipulated; an algorithm based on hashes, on the other hand, cannot.^[49] However, fingerprinting technologies cannot be used for all file formats. Reason being that the algorithm of eg an audio fingerprinting tool examines, among other things, the frequency values in a music file and thus cannot be used to identify copyrighted photos. If a technology is to examine all files on a web page, there would need to be a different fingerprinting tool for each type of media. Given that the range of copyrightable content is very broad, there is no fingerprinting tool for many types of content (eg, for architectural designs).^[50]
15. Content ID can be used for classic uploads as well as for live streams and thus in real time.^[51]
16. (bb) Possible results and accuracy of the algorithm. The following results are conceivable:^[52]

1. The content is identified as a copy, the identification is correct (‘true positive’).
2. The content is identified as a copy, the identification is false (‘false positive’).
3. The content is not identified as a copy, the identification is correct (‘true negative’).
4. The content is not identified as a copy, the identification is false (‘false negative’).

17. The accuracy of an algorithm is calculated with the sum of ‘true positives’ and ‘true negatives’ divided by the total number of evaluated contents.^[53] Thus, an accuracy of 99% means that 99% of all content is correctly identified. However, the value says nothing about how well the algorithm performs in terms of ‘false positives’ and ‘false negatives’, ie whether the remaining 1% of cases are more ‘false negatives’ (disadvantage for the rightsholder) or more ‘false positives’ (disadvantage for the uploader). The ability of an algorithm to discern false positives and false negatives can be found out by using other metrics.^[54]
18. According to Google, the algorithm works with 99,7% accuracy in the music section.^[55] Uploaded content can be blocked or monetized by Content ID within minutes up to hours of publication.^[56] Efforts to circumvent the filter, such as by rotating the image, slowing or speeding up the audio track, or changing speed, pitch, or sound quality, are also detected.^[57] No automated system, however, can evaluate content contextually, for example, to determine whether the ‘fair use’ principle is relevant.^[58]

1.1.1.2.2.3 Procedure in Case of Collision

19. In YouTube’s rights management system, the so-called assets can be administered.^[59] In individual cases, reference files may overlap or several rightsholders may claim content ID rights to a video.
20. In the former case, in the event of a reference file collision, the rightsholder who uploaded the ‘latest reference file’ is notified of the collision.^[60] Within 30 days, the latter can either ‘claim ownership rights’ or ‘exclude a reference overlap’.^[61] If both rightsholders insist on exclusive rights to their reference file, the claims remain with the partner who provided the reference file first.^[62]
21. If multiple Content ID claims from competing partners relate to a video (eg, one claim relates to the audio track, the other to the image), the most ‘restrictive’ legal consequence (blocking instead of mere monetization) is applied.^[63] By contrast, the legal consequence of observation intervenes when insufficient information is available for certain assets.^[64] If two different rightsholders have content ID claims to the same asset, the policy of the rightsholder who owns the rights in the country of upload is applied; if both partners are supposed to own an asset in the same country, the strongest legal consequence is applied again.^[65]

1.1.1.2.3 Other Applications of AI in the Context of the Content ID-Procedure

22. In the context of the Content ID process, artificial intelligence is also used to search for incorrect reference material.^[66] This is accompanied by human control, also with regard to the correct application of the Content ID system.^[67] Google does not disclose exactly how this control is carried out.

1.1.1.3 Effectiveness and Efficiency Potentials in the Use of AI

1.1.1.3.1 Detection of Rights Violations

23. Effectiveness and efficiency gains are the most significant benefits of using algorithms to detect copyright infringement. For example, according to a study by Gray/Suzor, content is blocked within ‘minutes to hours’ after it is first published.^[68] Without automatic matching against YouTube’s database, rightsholders would not be able to detect anywhere near as many copyright infringements in a correspondingly short period of time. With ongoing investments in the further development of the Content ID system,^[69] the platform operator promises to make the algorithmic law enforcement system technically more reliable and more accurate on an ongoing basis,^[70] but also to improve the complaint procedure (in the interest of users).^[71]

1.1.1.3.2 Costs

24. From the perspective of rightsholders, the Content ID procedure is a cost-effective way of (provisionally) enforcing their own rights, especially in comparison to state court proceedings.^[72] However, shifting the burden of complaint^[73] and cost to users makes them refrain from filing appeals against enforcement measures of the platforms before state courts – out of fear of high legal costs.^[74]

1.1.1.3.3 Flexibilization and Individualization of the Sanction Regime

25. The Content ID system is also characterized by a high degree of individualization. It enables rightsholders to use ‘guidelines’ to decide in advance exactly which legal consequence a Content ID claim should be directed at.^[75] Rightsholders can, for example, specify that Content ID claims addressing users from the USA are to be monetized, while videos are to be blocked if the uploader is located in Germany.^[76] It is also possible to allow short uploads by fans for promotional purposes.^[77]

1.1.2 Algorithmic Content Moderation on Communication Platforms

1.1.2.1 Terminology

26. The term ‘content moderation’ is not defined by law. In general, content moderation describes the process online platforms typically use (i) to set own standards of public discourse, information flow, and individual freedom of expression; (ii) to enforce those standards and (statutory) rights, particularly by means of screening, ranking, filtering, and blocking user-generated (unlawful) content, including proactive (and reactive) algorithmic techniques; and (iii) to establish internal proceedings with the primary purpose of protecting user rights more effectively.^[78] According to a broader definition, content moderation is meant to be ‘the governance mechanisms that structure participation in a community to facilitate cooperation and prevent abuse.’^[79]

1.1.2.2 Incentives for Establishing Ex Ante Mechanisms of AI-Based Content Moderation

27. From the perspective of online platforms, there are multiple reasons for establishing ex ante automated technology allowing to find, remove, and prevent illicit or standard-infringing user generated content. First and foremost, internet service providers are driven to install such means in order to avoid or, at any rate, minimize potential risks of liability.^[80] By the same token, platforms might be tempted to pre-empt potential regulatory constraints tightening (intermediate) liability, or even to bow to the pressure of powerful rightsholders to stem the flood of infringing content more effectively compared to a reactive notice and takedown system.^[81]
28. Even though there is no explicit statutory obligation for online platforms to actively seek facts indicating infringing user content – neither under the EU DSM Directive,^[82] the E-Commerce Directive,^[83] and the Digital Services Act^[84] nor under the US DMCA^[85] –, (large) Internet platforms actually developed corresponding monitoring and filtering techniques.^[86] Rather, the EU legislator^[87] and the CJEU^[88] tolerate the use of algorithmic filter technologies.^[89] On the contrary, intermediary immunities and safe harbors^[90] stimulated the proliferation of innovative moderation technologies.^[91] 
29. At least, developing such programs helps service providers to promote their own business and reputation while, at the same time, protecting their users.^[92] In this lens, platforms have, in the course of the (recent) past, undertaken substantial investments in techniques enabling them to proactively monitor or filter illegal and standard-violating online content.^[93]

1.1.2.3 The Functioning of Algorithmic Content Moderation

1.1.2.3.1 Basic Features

30. Content moderation can take place both automatically and by human moderators.^[94] In communication platforms, both variants are usually used (hybrid moderation system), albeit the proportion of automated content moderation varies considerably depending on the platform.
31. In hybrid moderation systems, automated content moderation has three main functions: (i) detecting violations of law or platform standards;^[95] (ii) supporting human content moderators, for example by prioritizing submitted content^[96] or by means of a preliminary assessment of the content in question;^[97] and finally (iii) independent decision-making by the moderation system.^[98] In addition, automated content moderation is meant to prevent the re-upload of content that violates the law or the terms and conditions of the platform (so-called ‘stay down’).^[99]
32. Automated content moderation can take place both proactively – through the use of filtering technology at the time between upload and publication of the content^[100] – and reactively after publication of the incriminated content on the platform (eg, in response to flagging of the content by users).^[101]

1.1.2.3.2 The Technical Operation, Inter Alia, Machine Learning Systems

33. Automated content moderation can be static – as in copyright law – in order to classify known content by means of fingerprinting (hash matching),^[102] as well as dynamic on the basis of machine learning tools: The goal here is to detect and sanction new, unknown content. In addition, there are hybrid systems that have elements of both static hashing/matching and dynamic classification methods.^[103]
34. Concrete technical tools for dynamic detection of incriminated content are primarily so-called classification tools, which use statistical methods to predict the probability of an infringement.^[104]
35. The decision-making process in machine-learning based on content moderation is highly dynamic: The artificial intelligence identifies certain patterns and correlations and adjusts its sanctioning reaction (allowing the upload, blocking or removal) accordingly.^[105] This can be done by labeling (‘offensive/not offensive’) as part of a supervised learning.^[106] In that respect, analysis techniques are applied to large amounts of data. Based on valid training data, the artificial intelligence is capable of learning, recognizing patterns and classifying on the basis of probabilities with regard to the installed labels. Subsequently, a sanction can be enforced automatically.^[107] Finally, the use of deep learning technologies with the aid of artificial neural networks is also conceivable.^[108]
36. Information about content identified as illegal is fed back into the machine learning system. Through such a feedback mechanism, the system ‘learns’; a so-called feedback loop is created.^[109]
37. Machine learning requires considerable amounts of data.^[110] On one thing, these are fed by the data on the platform, for example about user activities.^[111] Certainly, data available outside the specific platform can also be included, eg, by connecting third-party applications via application programming interfaces (APIs).^[112]
38. In the light of different objectives amalgamating in an AI system, machine-learning based content moderation finds application both to detect and sanction incriminated content and with the aim of commercializing content, eg, by personalized matching of this content.^[113]

1.1.2.4 Excursus: The Content Moderation of the Meta Group as a Prototype of Individualized Process Design^[114]

39. While originally, Meta’s content moderation was only rudimentarily regulated and based exclusively on unpublished guidelines,^[115] it henceforth draws on an increasingly complex construct of hierarchically structured private rules. These are enforced as part of a multi-track – now largely automated – internal dispute resolution process.^[116]

1.1.2.4.1 The Emergence of Specific Types of Platform Procedures Using the Example of the Cross-Check System

1.1.2.4.1.1 Procedural Framework

40. Meta even has its own internal procedures for certain groups of users (‘entitled entities’). Content posted by these persons, which is subjected to a review as part of the moderation process, is treated separately as part of the so-called cross-check system.^[117] In doing so, users are accorded a more prominent position in terms of the procedural and substantive legal framework.^[118] The specific procedural position of the user is in some cases largely determined by its economic relevance for the Group.^[119] This does not only apply in comparison to the regular platform procedure. Such a differentiation is also found within the cross-check system.
41. The procedure as such is tailored to the correction of so-called ‘false positives’ – ie content decisions that were wrongly classified as a violation of the in-house community standards during the regular moderation procedure, but are in fact to be judged as compliant.^[120] As a result, a cross-check procedure is only opened if the content in question is found to violate the community standards during the regular moderation procedure (‘at scale’). However, the opposite constellation of so-called ‘false negatives’ is not covered. This is content that violates the company's own standards but is incorrectly qualified as compliant during the regular moderation process.^[121] This means that the procedure is unilaterally geared towards preventing overenforcement.^[122] This is a procedural design that the Meta Group explicitly supports with regard to the content covered by the system, especially that of important advertising customers.^[123]
42. The cross-check procedure can only be opened by the Meta Group (quasi ‘ex officio’). Therefore, it is designed as an internal review procedure. Users have no possibility to initiate such a procedure (eg, by filing an in-house appeal).

1.1.2.4.1.2 Specifics of the Procedure

1.1.2.4.1.2.1 Context-Specific In-Depth Examination

43. Content is prioritized as part of the cross-check system and reviewed by particularly expert moderators located within the company (‘Early Response Secondary Review [ERSR]’). This review is always performed by human moderators. The moderators carry out an in-depth review of the content, considering the specific context of the statement, and have special decision-making powers. For example, at an advanced stage of the cross-check process, they can grant exceptions for content that is contrary to Facebook community standards. This is done by applying specific policies^[124] and exceptions^[125]. Both are not publicly accessible^[126] and can only be applied within the scope of such special procedures – ie, not in the regular procedural course (‘at scale’)^[127] – by selected bodies within the company.^[128] As a consequence, they are denied to the vast majority of Facebook users. The application of such specific internal regulations and exceptions conflicts with Meta’s statement that identical standards are used as the basis for content moderation on all platforms.

1.1.2.4.1.2.2 Suspension Effect

44. In terms of timing, the cross-check procedure is downstream of the regular moderation procedure (‘at-scale’). It takes place immediately after the original moderation decision (‘initial decision’), but before the intended private sanction is enforced. In this process, the content to be reviewed remains on the platform until the cross-check procedure is completed^[129] – at least within the scope of the person-related Early Response Secondary Review.^[130] From a structural perspective, initiating the cross-check procedure has a kind of suspension effect. This represents a significant difference from the regular moderation procedure and other special procedural steps of the moderation process. In this case, the sanction imposed in the initial decision remains enforced until the final decision, even in the event of a renewed review of the content. This applies both on the basis of an in-house appeal filed by the user and on the basis of an autonomous review initiated by the Meta Group.

1.1.2.4.1.3 Special Modalities of the Cross-Check Procedure

45. Within the cross-check procedure, a differentiation is made between a person-related (so-called Early Response Secondary Review) and a content-related procedure (General Secondary Review).

1.1.2.4.1.3.1 The Person-Related Strand of the Cross-Check Procedure: The ‘Early Response Secondary Review (ERSR)’

46. Content from users (‘entitled entities’) that violates community standards, which is on an internal list (‘cross-check list’) specifically intended for this purpose, is always subject to a follow-up check as part of the cross-check procedure (Early Response Secondary Review, in the following: ERSR).^[131] This particularly includes business partners of the Meta Group, advertising customers and users who represent a particular legal or regulatory risk for the Group – for example, in the case of ongoing legal disputes – as well as state actors.^[132] Also addressed are prominent groups of persons such as journalists.^[133] The qualification of a user as an ‘entitled entity’ is at the discretion of the Meta Group. According to the company’s own Oversight Board, economic parameters in particular are decisive for this.^[134]

1.1.2.4.1.3.2 The Content-Related Strand of the Cross-Check Procedure: The General Secondary Review (GSR)

47. In addition, a cross-check procedure can also be initiated for content-related reasons, irrespective of the identity of the posting party.^[135] This procedural modality – the so-called General Secondary Review, in the following: GSR – was only created retrospectively in 2021,^[136] presumably as a reaction of the group to the Facebook Files scandal.^[137]
48. The detection of such GSR content is carried out by an algorithm (‘cross-check-ranker’), resembling an automated triage process. In principle, all content on the platform is suitable, regardless of the identity of the posting user.^[138] The content in question must already have been classified by the regular moderation system as a violation of the community standards and be intended for enforcement of a corresponding platform sanction.^[139] Trivial is whether this is done by a human moderator or automatically.^[140]
49. The algorithm’s key decision-making criteria include the sensitivity of the post and the user concerned, the potential enforcement severity of a platform sanction, the probability of a false positive decision, and the potential spread of the content in question.^[141] The cross-check ranker continuously identifies new content suitable for GSR. As a result, previously unprocessed, older and lower-priority content is downgraded in the internal processing order of the GSR process and, due to the time limit of the process, eventually drops out of the GSR process altogether.^[142]

1.1.2.4.1.4 Allocation of Resources in Favour of the Personal ESRS Procedure

50. A gradation also takes place within the cross-check procedure. Content in the personal ERSR process is prioritized and given priority for review. In that regard, it is guaranteed that the content of the ‘entitled entities’ is checked.^[143] In contrast, in the content-based GSR procedure, content is only checked once there is any residual capacity:^[144] If there is no such capacity within a certain period,^[145] this GSR-related content falls outside the scope of the cross-check procedure. This ends the privilege of this content remaining on the platform during the ongoing procedure and the sanction of the regular (‘at scale’) procedure takes effect. This weighs heavily in particular due to the very high overturn rate in the GSR procedure: For example, in February 2022, 80% of the content that was classified as a violation of the community standards by the regular moderation system was classified as compliant with those standards in the GSR procedure.^[146]
51. Contrarily, there is no such time limit in the ERSR procedure: Here, the content remains on the platform until the final decision is made.^[147] As a result, content from users who are part of the ERSR system is always reviewed by a human moderator.
52. The allocation of resources in favor of the ERSR procedure also means that content from the GSR procedure reaches the higher instances of the cross-check procedure less frequently. However, since the specific policies and exceptions described above can only be granted at an advanced stage of the cross-check procedure, this means that the vast majority of GSR procedures do not benefit from them.^[148]

1.1.2.4.1.5 Availability of Appeal Mechanisms in the Cross-Check Procedure

53. The divergent availability of appeal mechanisms within Meta’s content moderation continues within the cross-check procedure. In that regard, there is only partial internal ‘legal protection’ against decisions made in the cross-check procedure.^[149] Firstly, this applies to legal remedies for the affected users themselves (the posting party). Consequently, third parties who wish to appeal against the content of a user who benefits from the cross-check procedure (such as an advertising customer) are denied internal legal protection.^[150]
54. Under both procedural modalities, there is no notification that the content (or user) in question is part of the cross-check procedure.^[151] Users who are part of the cross-check system also have no opportunity to defend themselves against inclusion in the special procedure.^[152]

1.1.2.4.2 Other Specific Types of Procedures

55. There are also other specific types of proceedings. However, considerably less public information exists about how they work. As far as can be seen, these procedures also pursue specific purposes, such as preventing overenforcement or enabling the Meta Corporation to react to acute political crises and content that is likely to reach a viral level on the platform.

1.1.2.4.2.1 The High Impact False Positive Override (HIPO)

56. The High Impact False Positive Override procedure (in the following: HIPO) is also used to correct false positives.^[153] Therefore, it pursues the same objective as the cross-check procedure. HIPO is also a verification procedure initiated by Facebook itself. Users do not have the option of initiating it. Contrary to the cross-check procedure, however, the opening of a HIPO procedure does not result in a suspension effect: The sanction imposed in the initial decision remains in place during the subsequent HIPO procedure; a review is therefore only carried out ex post.^[154]
57. The detection of HIPO-suitable content can be automated.^[155] A check is then performed as part of the separate HIPO procedure track. The order of the review is based on an automatically assigned score that depends on the priority of the content (‘HIPO ranker’). Decisive criteria for this prioritization are also the sensitivity of the content and the user, as well as the anticipated degree of dissemination of platform content.^[156] Meta’s Oversight Board criticized the functioning of this HIPO ranker as ‘highly inaccurate’.^[157]
58. The review in the HIPO procedure is carried out by external (out-sourced) moderators.^[158] A review in the HIPO procedure only takes place if sufficient personnel capacities are available. If this is not the case, there is no additional human review at all; the moderation decision is made automatically.^[159]
59. This structure is similar to the General Secondary Review (GSR) under the cross-check procedure.^[160] Nonetheless, it remains completely unclear according to which criteria content is proposed for review in the GSR procedure and when merely a HIPO procedure takes place. This classification has significant consequences: For example, the content in question remains on the platform during the GSR procedure, while a HIPO review only enables a downstream review.^[161] It can probably be assumed that content deemed less relevant by the Meta Group will only be dealt with subsidiarity in the context of the HIPO review.

1.1.2.4.2.2 Procedures for Responding to Political Crises and Viral Content

60. The Meta Group provides an option for subjecting potentially viral content to additional human review (so-called High-Risk Early Review Operations: HERO).^[162] The same internal team that also carries out checks as part of the cross-check procedure is responsible for this.^[163] However, it remains unclear for which procedure the personnel resources are allocated. In any case, it seems likely that the person-related ERSR procedure will be prioritized.

1.1.2.4.2.3 Procedure for Handling Government Requests

61. The so-called escalation process provides for a separate procedure for reports by prosecution agencies.^[164] This applies both to reports based on potential illegality or relevance to criminal law, but also in particular to reports based on alleged violations of Meta’s own community standards.^[165] The sanctioning of content reported on the basis of a potential law infringement is regularly only carried out locally,^[166] for example through geo-blocking.^[167] In contrast, content that violates the community standards is blocked globally.
62. In terms of its objective, this procedure is to a certain extent the ‘mirror image’ of the cross-check procedure: In the government request procedure, the identity of the reporter is decisive, whereas in the context of the cross-check procedure, the identity of the posting party is the relevant criterion. In the escalation procedure, an in-depth review of the content is also carried out by internal company moderators.^[168] Likewise, there is no possibility of redress against these decisions:^[169] Users whose content is sanctioned by government agencies on the basis of a report are thus denied legal protection within the company solely on the basis of the special status of the complainant. The situation is further aggravated by the fact that the user concerned is not informed of this fact either.^[170] The escalation procedure also opens up the possibility of applying specific policies and exceptions.^[171] Furthermore, government agencies are not required to substantiate their reports – for example, by naming the community standard presumed to be affected, by justifying the enforcement decision or by providing a sufficient basis of evidence.^[172]
63. The procedure for reporting by government agencies always takes precedence over the cross-check procedure. In this case, the company’s internal decision-makers are in a position to make sanction decisions directly – irrespective of the specific content or the identity of the posting party. This applies to both modalities of the cross-check procedure, in particular also to ‘entitled entities’ under the ERSR procedure.^[173]
64. Overall, sovereign actors are thus privileged in two ways. In their favour, separate procedures apply both to content posted by them and to content reported by them.

1.1.2.4.3 Conclusion

65. Meta’s internal procedures show a progressive automation of content moderation. In that regard, rapid technical progress offers the possibility of ever greater flexibility and individualization of platform procedures.^[174] Their design is apparently based on economic interests: namely a preference for advertising and business partners, the minimization of liability risks and the avoidance of negative public perception. This is particularly clear from the privileged treatment of user groups in the ERSR procedure and state actors in the government request procedure. Moreover, the design of the internal platform procedures focuses on preventing overenforcement. In contrast, Meta’s efforts to prevent underenforcement^[175] are only weakly developed – probably to avoid the impression of censorship.^[176] 
66. The possibility of a company-internal user appeal depends on the discretion of the platform. Its predominant role becomes particularly apparent in cases of purely automated decisions (and technical deficits in the system), not least if a user has no other effective legal protection. In view of the fact that platform-based dispute resolution substitutes traditional legal protection, the design of internal company legal remedies should not be subject solely to the private design power of platforms, but – as recently in the DSA (esp in Art 20) – should be accompanied by an effective guarantee of basic procedural safeguards in platform proceedings and corrected where necessary.

1.2 Dangers and Drawbacks of AI-Based Law and Standards Enforcement

67. An AI-based enforcement of private rights is often viewed critically. Using Content ID as an example, the following section describes the main dangers and disadvantages that can arise from a computer-based (copy-)rights check or from the private enforcement of (copy-)rights. Legal issues arising solely from the ‘regulatory framework’ of the Content ID process will be excluded.

1.2.1 The Economization of Intra-Platform Decisions: Manifestations and Technical Design of Platform Power

1.2.1.1 Structuring of Interactions on Platforms

68. As an integral and dominant part of the market, digital platforms enable the exchange of goods and services, the processing of payments, and global communication by means of their technical and economic infrastructure.^[177] The platform economy is essentially based on the use of continuously refined algorithms.^[178] In addition to providing and structuring the infrastructure, platforms as economic actors primarily pursue their own economic interests – whether through participation in generated profits,^[179] by placing advertisements^[180] or also for the purpose of avoiding liability. Both functions of the platform merge in a technical system.^[181]
69. As a consequence of the law-enforcing and dispute-settling role that platforms are increasingly assuming, classic procedural guarantees – such as the right to be heard – are being dosed on the basis of economic parameters, ie, limited or even denied comprehensively.^[182] This is driven in particular by tendencies toward privatization of tasks traditionally incumbent on state actors.^[183]
70. The agglomeration of users and the tendency to form monopolies is inherent in the economy of platforms.^[184] Platforms structure their markets (explicitly) and act as private rule-setters.^[185] For example, e-commerce platforms create digital market orders and structure the exchange of goods between their customers;^[186] communication platforms act on the communication of their users. In the context of content moderation, social media platforms set their own (behavioural) standards and implement them on the basis of their own sanctioning system. In doing so, they act functionally comparable to a court.^[187] The structuring of interactions on platforms also – if not primarily – takes place implicitly through the architecture, the code of the platform.^[188] In technical terms, this behaviour control is highly individualized and – through the use of algorithms and machine learning tools^[189] – adaptive (so-called hyper nudging).^[190] By means of behavioral microtargeting,^[191] it is possible to address users in a personalized manner on the basis of behavior and personality-based user profiles. These profiles are generated by algorithmic analysis of large volumes of data. In particular, the user’s behaviour on the platform is considered.

1.2.1.2 Tailored Outcomes and Procedures

71. The alignment of the platform infrastructure based on economic parameters takes place at the level of access to the procedure and throughout all stages of the procedure. This makes it possible to tailor both the concrete design of the procedure and the substantive (enforcement) decision to the individual user.^[192]

1.2.1.2.1 Individualized Process Design in Payment Transactions

72. In the case of payment service providers, it is known that an individually created score for the user, which is based on a user’s economic profitability, for example, determines the level of expertise of the decision-maker that the platform uses to moderate or settle a dispute in each case.^[193] The material content of decisions – eg, when granting goodwill – is also based individually on the ‘value’, ie, the economic relevance of a customer. Admittedly, this is not a new development. However, the possibilities for analyzing large volumes of data, which have been considerably increased by means of artificial intelligence, are making it easier to examine and evaluate customer behaviour and to react to it (in a timely manner).^[194]

1.2.1.2.2 The Content Moderation of the Meta Group as a Prototype of Individualized Process Design

73. The development outlined becomes particularly vivid around communication platforms. The Meta Group serves as the inspiration for this. It has been known for some time that in the moderation process of Facebook’s own platform,^[195] the factors of virality, the potential danger of the content and the probability of a violation of the company’s proper community standards determine whether decisions are made automatically by an AI system or by a human content moderator.^[196] In particular, this development can be observed on the basis of the Group’s special platform procedural traits. The impression suggests itself that the establishment and design of these procedures primarily serves the pursuit of economic interests – for example through the preferential treatment of advertising and business partners –, furthermore the minimization of liability risks and the avoidance of negative public perception. This becomes evident in the privileged treatment of certain user groups under the cross-check procedure and the provision of a specific procedure for notifications by government agencies.^[197]

1.2.1.3 Private Ordering and Monetization in Copyright Law

74. The alignment of the platform infrastructure with economic self-interest is not less vivid in the area of copyright, as is exemplified by YouTube’s content ID module.^[198] With the U.S. DMCA, the DSM Directive of the European Union, and most recently the German Copyright Service Provider Act (UrhDaG), there is (by now) a narrow regulatory framework for dispute resolution – compared to the law of expression. Nevertheless, there is a clear shift from the classic notice-and-takedown system to the monetization of content.^[199] Primary beneficiaries, as shown, are platforms and rightsholders.^[200] The design of Content-ID as well as technically continuously optimized (upload) filter technologies have a structural effect to the detriment of the uploader (encroachment on freedom of expression,^[201] danger of overblocking or chilling effects).^[202] Last but not least, the obligation to use filter technology opens up new markets and thus investment incentives in the economic interest of platforms and software manufacturers to develop increasingly precise filter technologies.^[203] The shift toward monetization shows that platforms know how to use the regulatory leeway provided by the state to their own economic advantage. Corresponding technological developments are based on state regulations, insofar as they have a mandatory character. Beyond this, they still often deviate from the legal model in crucial areas and become independent. YouTube’s content ID is paradigmatic for the further development of private law enforcement and dispute resolution systems from state guidelines and principles of copyright law (such as uniform law enforcement and the application of exemptions) to complex systems of private orders.^[204]

1.2.1.4 AI-Driven Structuring of Communication Processes and Separation Of User Groups

75. The social, political and private dimensions of platform use cannot always be clearly separated from the underlying commercial transaction, especially on social media platforms.^[205] Here, the moderation of harmful or even (criminally) illegal content merges with the commercial goal of encouraging users to stay on the platform as long as possible in order to collect data and expose personalized advertisements.^[206]
76. The described mixed situation is reinforced by the technical functioning of algorithms: In particular, the concept of ‘filter bubbles’ assumes that algorithms, due to their technical mode of operation, tend to display to the user only information that matches his or her previous views and to hide information that does not fit.^[207] The phenomenon of algorithmic segmentation of groups of people into so-called echo chambers – which is quite controversial in its scope – is also discussed.^[208] In this context, Internet users interact only with like-minded people. Such separation harbours the risk of intensifying discriminatory ideas and radicalizing communication processes (‘algorithmic radicalization’).^[209] Adaptive platform environments and the use of feedback loops in turn encourage user behavior influenced in this way to be ‘fed back’ into the system. This shows: The configuration of the algorithm determines which content and information is displayed.^[210] In doing so, the algorithmically controlled architecture of platforms has a very significant influence on the freedom to form and express opinions, as well as on the guarantee of media pluralism as a whole.^[211]
77. Algorithmic separation of certain groups of people also takes place on platforms with a direct link to the market,^[212] especially in e-commerce^[213] and on sharing platforms^[214]. However, the encroachment on areas sensitive to fundamental rights is less pronounced here.

1.2.2 Merging Government and Private Regulatory Objectives: The Public-Private Divide

1.2.2.1 Privatization of Law Enforcement and Dispute Resolution

78. In the area of law enforcement and dispute resolution in particular, it is becoming apparent that platforms are increasingly performing tasks that have traditionally been the responsibility of state actors,^[215] including the judiciary.^[216] Complex privatization processes are emerging across all regulatory and legal areas – whether in copyright law, the law of expression, data protection law^[217] or the settlement of contractual disputes in e-commerce or payment transactions.
79. The causes of this development are intricate. Some of them stem from complex governance problems, especially in the case of cross-border conflicts.^[218] At the same time, state regulation is shifting dispute resolution to private actors:
80. In copyright law, such a shift of responsibility^[219] manifests itself, for example, in the imposition of procedural obligations,^[220] which in turn are implemented and (excessively) concretized by private forms of dispute resolution such as YouTube’s Content ID.^[221] The DSM Directive sets incentives, if not a de facto obligation, to use filtering systems.^[222]
81. In the area of the law of expression, platforms regulate the boundaries of what can be said as part of their content moderation by means of private standard-setting and technical infrastructure^[223] – even beyond the area of illegal content and content relevant to criminal law.^[224] This development is also driven by economic motives, such as maintaining contractual relationships with advertising customers. For example, in response to criticism from major advertising customers, Google announced that it would refine the content classifiers it uses to detect extremist and terrorist content.^[225] This reaction was preceded by the placement of ads on websites with antisemitic and homophobic content.^[226]
82. As the German Network Enforcement Act (Netzwerkdurchsetzungsgesetz, in the following: NetzDG) exemplifies, state regulation shifts the enforcement of state law to private parties.^[227] Where legal leeway remains, the German legislator explicitly relies on self-regulation^[228] and accepts that platforms transfer their (primarily automated) dispute resolution systems, which they have developed for complaints against their general terms and conditions, to the enforcement of state law – ie, matters within the scope of the NetzDG.^[229] The concrete modes of control are manifold: In addition to procedural rules, the NetzDG also relies on indirect behavioral control by means of liability obligations^[230] and the threat of fines.^[231]

1.2.2.2 (Cross-Border) Influence of State Actors on Platforms

83. A reflection of this is a growing state influence on globally effective private regulatory systems of large online platforms. By exerting regulatory impact on the governance structures of such platforms, states are able – whether intentionally or unintentionally – to achieve a global effect of their regulatory acts (which in themselves only have a national impact).^[232] The main reason for this is the position of digital platforms as gatekeepers for key social and economic needs. By adapting their technical infrastructure to the regulatory requirements (of the European Union, for example), platforms are in a position to help these requirements achieve transnational, even global validity.^[233]
84. The CJEU’s decision in Glawischnig-Piesczek v Facebook serves as a model for the de facto extension of national law. The ruling approved both the worldwide blocking of the incriminated content and the obligation of the platform to prevent future infringements with content similar in wording and meaning by the Austrian court of instance. The CJEU explicitly stated that hosting operators are allowed to use ‘automated techniques and means of investigation’ – such as keyword filtering^[234] – and are therefore not obliged to make ‘autonomous assessments’.^[235]
85. Such ‘regulatory reinforcement’ can take place explicitly, for example in the form of legislative acts. Increasingly, however, it is also possible to observe the exertion of informal pressure.^[236] This ‘jawboning’^[237] is expressed, for example, in the form of threats by political actors or the activity of so-called Internet Referral Units (IRUs).^[238] These entities are state agencies that report content to the service providers on the basis of violations of the platform’s own standards – and thus not on the basis of violations of state law – and encourage removal, usually on a global basis.^[239] Platforms are not obligated to remedy the deletion requests and remove the content; rather, they act voluntarily.^[240] Nevertheless, these requests represent a vehicle of state influence: For example, mass and disproportionate reporting of specific content risks imposing a state-preformed understanding of platform standards on platforms.^[241] The often broad and open wording of these standards benefits state actors in this regard.^[242]
86. Moreover, the service providers are quite cooperative: The Meta Group, for instance, prioritizes the processing of inquiries from such internet referral units within the framework of a separate, so-called government request procedure, with the involvement of expert decision-makers equipped with special competencies.^[243] In addition, there is no internal legal protection against moderation decisions made in response to a report from government agencies.^[244] Although such reports are also reviewed exclusively due to a company’s own community standards, it is possible to observe a shortening of the company’s internal legal protection simply because the complainant is a sovereign actor. YouTube also offered to classify the Internet Referral Unit of the EU as a Trusted Flagger.^[245]
87. By participating in industry-specific hash databases of service providers, state internet referral units also explicitly exercise ‘definitional power’.^[246] In that regard, the regulation of platforms varies from country to country: The United States, Great Britain and Germany are considered particularly influential.^[247]
88. Given that the entire content moderation of a platform takes place within the same technical system, the adaptation of state regulation has resulted in an increasing fusion of two – functionally fundamentally different – forms of enforcement: the enforcement of state regulations in the guise of private, platform-specific standards. The technical peculiarities of machine learning used in content moderation further encourage this development.^[248]
89. In copyright law, too, the obligations of the U.S. DMCA are being implemented on a de facto global scale by incorporating the relevant regulations into the technical infrastructure of the platform. Accordingly, blocking because of DMCA notices often takes place globally and not in the form of geo-blocking limited to local effects.^[249]

1.2.3 Technical Limitations and Restricted Traceability of Context-Dependent Content

90. AI systems are subject to technical limitations: While AI-based upload filters already work reliably when it comes to identical sound and image files, the result for the recognition of static image files in ‘dynamic audiovisual files’ is (still) significantly less favorable.^[250]
91. A central point of criticism is furthermore the difficult tangibility of context specific, time-sensitive and locally dependent content by AI.^[251] Systems based on the fingerprinting method have – at least so far^[252] – difficulties in distinguishing copyright-legitimate from certain unlawful reproductions, primarily regarding the scope of application of the fair use exception under US copyright law.^[253] The same applies to typically context-dependent content such as criticism, caricatures, parodies, pastiche, or quotation.^[254]
92. Considering this, parts of the literature assume that algorithms are generally not capable of legally adequate assessments and scrutiny.^[255] The deficient recognizability of permitted content is seen as problematic, especially in view of a ‘remix culture’ whose creative approach consists precisely in alienating already existing content and protected works (see, for example, memes, supercuts or mash-ups).^[256]
93. Particularly in view of the limited algorithmic recognizability of fair use exceptions, the fundamental question arises as to whether a system that learns on the basis of comparative materials will be able to replicate abstract legal and statutory situations or, on the basis of case law, selectively decided legal questions of the past. Also questionable is whether a correspondingly static ‘law recognition and application program’ will be able to either subsume unknown facts (not yet fed into the system) under applicable, context-based legal provisions in accordance with the principles of legal methodology, further to develop the law accordingly, or ‘simply’ to represent dynamically developing legal situations in an accurate manner.^[257] This explains why fingerprinting and hashing technologies are quite effective modes of public-private cooperation on policing content that is predetermined to be unlawful, while they are unlikely to helpfully address highly context-dependent questions.^[258]
94. Finally, as far as can be seen, there is no adaptation to national (copyright) law, including the fair use exception, within the algorithm either.^[259] This aspect is reinforced by an ‘Americanization’ of the Content ID procedure given that YouTube^[260] seems to apply the US principle of fair use de facto worldwide.^[261]

1.2.4 Limited Transparency

95. Another point of criticism concerns the lack of transparency of AI-based law enforcement.^[262] In general, algorithms that are not disclosed or cannot be traced hamper creative user activity^[263] and generate incentives to circumvent the algorithm.^[264]
96. Moreover, there is a serious risk of bias with respect to the ‘preselected’ data, but also with respect to the algorithm itself.^[265] It is well known that even state judges are sometimes guided by arbitrary motives and dependencies. However, the decisive factor in state civil proceedings is whether a judicial decision, on account of its justification, is comprehensible in terms of its content: If this is the case, the actual motives of a judge are irrelevant as long as they do not give rise to any concern of bias or partiality. However, it is precisely this comprehensibility that is in question in the case of AI-based law enforcement:^[266] According to the so-called black box problem, the developers of an AI regularly do not themselves know or understand extensively how the algorithm works in detail,^[267] nor can they predict or subsequently provide information about what an AI has actually learned^[268] – this circumstance makes it considerably more difficult to deal with the AI and to disclose the internal processes.^[269]
97. This finding is reinforced not only by the fact that YouTube leaves the decision on the dispute largely to the parties or, primarily, to the rightsholder.^[270] In addition, since the actions of the online platforms are structurally guided by interests and liability,^[271] it is hardly possible to reliably act as a dispute moderator in a, in principle, independent manner.
98. Finally, non-transparent AI-based legal enforcement makes it difficult for the uploader to find out whether notifications were sent willfully or without sufficient legal research, thus weakening his effective defense options.^[272]

1.2.5 Overblocking

99. Technical limitations of AI, such as in particular a limited ability to map context-dependent content, can, in principle, lead to both underenforcement^[273] and overenforcement in the context of private law enforcement. According to the predominant assessment of the Content ID procedure, the limited recognizability of context-dependent permissions leads to structural overblocking, ie, to the sanctioning of user content that is wrongly recognized as illegal.^[274]
100. If an algorithm detects a copyright infringement even if only quantitatively minor parts are affected,^[275] and if the platform operator (as YouTube probably does in the context of the Content ID procedure^[276]) sanctions this infringement by the (alleged) infringer losing the possibility of monetization comprehensively to the (alleged) rightsholder, such an incentive reinforces the phenomenon of overblocking.^[277] It is obvious that in these cases the legal sphere of platform users is encroached upon, first and foremost their right to freedom of expression, thereby silencing legitimate or marginalized speech.^[278]
101. Not all categories are equally affected by content blocking: For instance, rightsholders decide significantly more often to block the incriminated content than to monetize it, especially in the case of illegal interference with exploitation rights to films or to content in the field of sports. In contrast, content ID claims relating to video games are enforced less frequently.^[279]
102. On an overall basis, YouTube is extremely cautious in taking action against abusive Content ID claims. In that regard, the company merely reserves the right to take ‘appropriate measures’^[280] and refers to its ‘Best Practices for Claims’.^[281] Whether YouTube checks references from rightsholders in advance is not clear.^[282]

1.2.6 Algorithmic Discrimination

103. Platforms are non-neutral decision-makers due to the pursuit of their own economic interests.^[283] This problem is reinforced and perpetuated by the use of AI-based decision-making systems. In the field of communication platforms, algorithms by means of which content is curated and prioritized are said to contain stereotypical and discriminatory assumptions.^[284] In any case, this encourages discrimination against marginalized groups and minorities: For example, expressions frequently used by English speaking African-American are flagged significantly more often; posts by Muslim users have disproportionately often been blocked in the past on the grounds that they are terrorist content.^[285] In contrast, the major platforms Facebook and YouTube have a significantly more lenient moderation practice with regard to racist and right-wing terrorist content, ignoring such content to a greater extent.^[286]
104. Algorithmic discrimination is usually not a consequence of an (unknowing or even conscious) implementation on the part of the programmers, but often occurs through the interaction of the AI system with the technical environment and the available data.^[287] Such biases can occur both at the level of the training process – for example, through the use of incomplete or unrepresentative data – and at the classification level, for example, by linking to forbidden distinguishing characteristics such as race or gender.^[288] Especially in AI-driven content moderation, the complicating factor is that a platform environment tailored to users feeds interactions of a user with this environment back into the system in the form of so-called feedback loops.^[289]
105. Due to the possibilities of individualizing procedures,^[290] discrimination takes place in all described procedural stages and thus also at the level of the decision itself. The problem of algorithmic discrimination is intensified by the lack of transparency of internal company decisions. In the case of AI-based moderation, this is further perpetuated by the opacity of the AI system itself (‘black box problem’).^[291]

1.2.7 Information Gaps and Insufficient Stakes

106. Given the opaque nature of AI systems and a lack of mandatory publication requirements pertaining to impact-assessment, users often lack insight into the way an AI-based content moderation system operates, but also in terms of whether a specific removal or other enforcement measure has affected them.^[292]
107. Further, in cases of legitimate speech being erroneously removed or blocked, the perceived harm might seem too small for a speaker to act upon and challenge the enforcement decision. In doing so, a platform operator does not solely cause harm to the individual user concerned but equally to the public which might be deprived of parts of the public discourse and of access to information.^[293]

1.3 Procedural Answers

1.3.1 The Need for (Complementary) Procedural Resolution Schemes

1.3.1.1 Initial Situation: Proceduralization of Private Enforcement

108. Internet platforms appear to be natural enforcers of rights: Their technical dominance over access to and use of networks (‘modern public square’^[294]) predestines them for the role of enforcing private rights – their own and those of others. This happens every day in a myriad of cases: In the period between April and June 2023 alone, Facebook sanctioned 8.9 million cases of alleged violations of its community standards on physical abuse and sexual exploitation of children worldwide;^[295] add enforcement actions for violations of legally conferred subjective rights or even specific norms of criminal law.^[296]
109. The intermediaries are involved in a process of clarification and evaluation of unclear factual and legal position. The German Federal Court of Justice, which recently had to rule in two high-profile decisions^[297] on whether Facebook’s deletion of suspected hate comments was lawful, even speaks of a binding procedural law that platforms with market power would have to establish in their terms and conditions as they are bound by fundamental rights due to their structural superiority.^[298] Under the topos of effective ‘protection of fundamental rights through procedure’,^[299] the Federal Court of Justice developed individual procedural requirements in order to collaterally give effect to the fundamental rights of platform users, in particular their fundamental right to freedom of expression, Art 5(1) Grundgesetz (Fundamental Law, hereinafter: GG), even in the context of private legal relationships.^[300]
110. The EU legislator is also vigorously driving forward the proceduralization of private enforcement. He is doing so by establishing procedures for the enforcement of rights and platform standards (general terms and conditions) as well as internal complaints, thereby assigning the role of ‘objective’ dispute resolvers to platforms: for example, in 2019 in the DSM Directive on the enforcement of copyrights in the digital single market,^[301] in the P2B Regulation^[302] or more recently, across legal areas, in the Digital Services Act of the European Union.^[303] In the scope of the E-Commerce Directive,^[304] the CJEU even recognized in the Glawischnig-Piesczek case the possibility for Member State courts to order hosting providers such as Facebook to delete or block illegal content worldwide.^[305]
111. This trend can most recently be observed in the regulatory strategy of U.S. law: The draft Platform Accountability and Consumer Transparency Act (‘PACT Act’), for instance, also requires the establishment of a complaint system for users, processing deadlines for incoming notices, information and justification obligations after a platform sanction has been imposed, and the publication of transparency reports.^[306] Texas and Florida have already passed similarly structured laws.^[307]

1.3.1.2 Emergence and Justification of Platform-Based Law Enforcement through Procedure

1.3.1.2.1 Liability Law and Basic Procedural Structures

112. The trigger for the emergence of procedural structures is the liability – indirect or perpetrator – of Internet platforms, for example, for copyright, trademark and personal rights infringements and for violations of the law of fair dealing by third parties:^[308] In this context, verification obligations, at least outside copyright law (DSM Directive), do not arise unsolicited – § 7(2) of the German Telemedia Act,^[309] Art 14(1), 15(1) E-Commerce Directive^[310] –, but in principle only as soon as the platform operator is made aware of a – regularly clear^[311] – infringement (so-called notice-and-takedown procedure).^[312] As a consequence, there is no general monitoring obligation:^[313] The right holder is burdened with a kind of obligation to present and substantiate^[314] the legal facts of the infringement and its factual basis in order to identify the legal position to be enforced^[315] and to trigger an obligation on the part of the intermediary to respond.^[316] The platform operators are, in turn, required to cooperate seriously in clarifying the facts.

1.3.1.2.2 Reasons for Platform-Based Law Enforcement through Procedures

113. The basic task of private enforcement of rights through proceedings is to (provisionally) reconcile legal positions that conflict with each other in an uncertain legal and factual situation – in the interest of the alleged infringer and the presumptive rightsholder alike. This individual-protective function of proceduralization is collectivized by scaling processes.^[317]

1.3.1.2.2.1 Structural Inferiority of the Addressee of Legal Enforcement: Empowerment of Effective Legal Protection through Procedure

114. Functionalizing Internet platforms for the enforcement of private rights decisively improves the enforcement chances of affected rightsholders, especially in procedural terms. The following aspects are crucial in this respect: Under the conditions of the Internet, state courts are often unable to enforce private rights within a reasonable period of time – even in case of provisional legal protection – or to cope with the systemic task of handling the uncountable number of cases.^[318] In addition, rigorous ex ante due diligence and verification obligations of platforms create additional incentives for excessive enforcement of private rights (overenforcement).^[319]
115. Such effects are amplified by an innate structural information advantage of the online platform over its users. Information asymmetries can already have an effect in the run-up to a sanction or legal enforcement measure and thus create strategic incentives for platforms to prevent the emergence of a legal dispute or the use of legal protection from the outset. Thereby making it more difficult, if not de facto impossible, for affected users to access internal complaints procedures or judicial legal protection.^[320]
116. In view of these initial conditions, there is a serious risk of a loss of effective legal protection. Without further procedural safeguards, platform users would be severely limited in their legal power to defend themselves effectively in and out of state courts and structurally inferior to the enforcement power of the rightsholder concerned. This is particularly true in view of the gatekeeper role large online intermediaries take over in participation in social life or in economic transactions – often in particularly sensitive areas such as access to housing, credit or jobs.^[321]

1.3.1.2.2.2 Fast and Effective Protection against Irreversible Damage Caused by Procedures

117. However, there are also mirror-inverted legal enforcement deficits: It is well known that the realization of rights on the Internet is particularly at risk. The reason for this is their disproportionately greater exposure as well as a faster and wider spread of infringements.^[322] This is accompanied by the difficulty of preventing infringements or reversing infringements that have already occurred.^[323]

1.3.1.2.2.3 Minimizing Erroneous Decisions: Procedural Law as an Instrument of Fact-Finding

118. Finally, effective legal protection is realized upstream in minimizing the risk of a factual or legal misjudgment of the enforcement decision, despite its normative provisional nature. In this context, it will be necessary to ask how procedural elements of factual clarification can be made usable for the platform operator to improve the decision-making basis of private legal enforcement.

1.3.1.2.2.4 Conclusion

119. Given the initial situation outlined above, especially the strong de facto absorption effect of platform-based legal enforcement vis-à-vis state legal protection, and in view of the inadequate protective and interest-balancing effects of a purely liability-based integration of platforms, it seems necessary to integrate the latter into effective legal protection through procedural structures and proceedings.
120. Considering the dangers of algorithmic platform behavior already described, this finding also applies to the enforcement of the platform’s own community standards. In the following, it will be shown whether and to what extent procedural structures go beyond guaranteeing legal protection for the individual and for a large number of individual affected parties.

1.3.2 Shaping Effective Legal Protection through Procedures: Potential and Limits of a Proceduralization of Platform-Based Legal Enforcement

1.3.2.1 Functions of Procedural Structures

121. The core task of corresponding procedural structures is to react quickly to particular risk situations, to create trustworthy (transparent) and error-minimizing decision-making conditions for this purpose, and to ensure the provisional nature (and correctability) of enforcement measures. In addition, there is a need for procedural structures capable of handling the mass of platform users – as potential violators of rights and terms and conditions as well as those affected by such violations.
122. In this orientation, proceduralization can strengthen the legitimacy of private law enforcement overall.^[324] Corresponding gains in trust, in the form of strengthened user loyalty, often also have an economic impact in favor of platforms.^[325]

1.3.2.2 Individual and Collective Law Dimensions of Effective Legal Protection

123. In view of the dangers (and potential) of algorithmic enforcement of rights and terms and conditions, the question arises as to which procedural instruments and guarantees can be used to adequately implement the basic tasks assigned to a platform procedure. Normative reference points from state civil proceedings are the requirements of effective legal protection and transparent procedures, which in principle can be applied to online platforms.
124. In its individual-protection dimension, the right to effective legal protection should apply regardless of whether platforms implement state-imposed rights of third parties or their own community standards. Thus, the Digital Services Act also recently requires platforms to take into account the fundamental rights guaranteed by the EU Charter in the context of both forms of enforcement, including the right to an effective legal remedy.^[326] As will be explained in more detail below, this requirement for legal protection should be shaped by various guarantees related to the platform. Particularly by effective access to complaints procedures (both internal and external and open to decision-making) and to a human decision-maker,^[327] furthermore by a guarantee of a fair hearing and procedural equality of arms and opportunities^[328] as well as by protection of the rights of the defense.^[329] Moreover, enforcement addressees must be protected against unjustified reports and the improper use of platform-based enforcement actions.^[330] In this respect, there are functional and structural parallels to state interim legal protection.^[331]
125. The multitude of potentially parallel, similar or even identical situations of impairment in relation to legitimate interests and rights of users and other affected persons, which are regularly caused by platform structures, also requires effective legal protection to be granted in a supra-individual, ie ‘collective’ dimension. In this context, procedural instruments and standards must be developed which, on the one hand, enable rapid and effective legal protection for a large number of affected persons and, on the other, provide procedural safeguards against the regulatory actions of platforms – for example, in the form of the scaling of interpretation standards or decisions.^[332]
126. Under EU law, elements of private self-monitoring and third-party control accompany the procedural monitoring mechanisms – in the form of company compliance departments to be set up^[333] or an annual independent audit.^[334] In addition, certain violations of the provisions of the Digital Services Act can be sanctioned with fines.^[335] This results in an overall multilayered regulatory arrangement.^[336]

1.3.2.3 Intra-Company Legal Protection Proceedings: Basic Structures and Procedural Guarantees

127. In light of this, it is firstly important to provide the enforcement addressee with an effective, ie easily and for all affected parties accessible (ie transparent and user-friendly^[337]), incidentally free of charge and fast^[338] legal protection procedure.^[339] This procedure, to be set up by the platform itself and designed for a large number of complaints, is intended to enable the enforcement addressee and affected right holders to effectively assert their own rights that are presumably affected by the enforcement of rights (enabling function of procedures). Governmental (or supranational) regulation has prescribed the establishment of corresponding procedures in various forms over the course of the past few years: for example, in the European Union through the Digital Services Act – particularly in its Art 20^[340] – or the P2B Regulation,^[341] but also in the USA^[342] and in Germany^[343].
128. In the following, individual procedural guarantees for the implementation of effective in-house legal protection are discussed in more detail below – detached from any previous affirmative action.

1.3.2.3.1 Online Platforms and Attributions of their Procedural Function

129. Various procedural models are available to reduce erroneous enforcement decisions and ensure the realization of rights in endangered situations. They vary, depending on the role or corresponding obligations that can be attributed to online platforms (and correspondingly to users and rightsholders or whistleblowers) in clarifying the facts relevant to enforcement.
130. According to the concept of ‘clarification responsibility’^[344], platforms act as de facto dispute mediators (Streitmittler). Hereafter, a host provider should be regularly obligated to forward the complaint of the affected rightsholder to the responsible party for comment before deleting an allegedly infringing blog entry. If a substantiated counterstatement raises justified doubts about an infringement, the affected party must in turn be given the opportunity to respond and, if necessary, be ordered to submit further required evidence.^[345] A corresponding ‘shuttle procedure’ makes the parties in the primary infringement relationship more responsible for clarifying the facts^[346] – such as the counternotification procedure under § 3b NetzDG (Network Enforcement Act).^[347] At the same time, the moderate model of the Federal Court of Justice borrows from a premise of the principle of submission in civil proceedings (Beibringungsgrundsatz), namely that the opposing interests of the parties grant the correctness of congruent statements of fact to a greater extent.
131. Nevertheless, this model reveals striking weaknesses: On the one hand, it leads to an expansion of the factual material relevant to enforcement through regularly disputed party submissions. This makes the clarification of the facts more complex and costlier, without online platforms being provided with the clarification tools of state civil proceedings or judges,^[348] in order to reliably clear up any ambiguities that often remain even after the parties to the dispute have been heard. Furthermore, the shuttle model – which is time-consuming in individual cases – is not sufficiently tailored to the urgent enforcement needs of affected, known and, above all, unknown rightsholders, nor to a large number of similar enforcement situations.^[349] Above all, however, it forces – especially in the case of the highly context-dependent facts of an infringement of personality rights, copyright or trademark rights^[350] – a weighing of conflicting legal positions in individual cases that is both time-consuming and prone to error, and thus leaves the forecasting risk for the actual existence of ‘clear’ infringements with the platform operator.^[351]
132. These disadvantages could be partially countered by further reducing or standardizing the inspection obligations of platforms – (also) in the interest of their increased predictability and fulfillment (‘model of procedural responsibility for action’^[352]): On the one hand, with regard to the substantive standard of review, in the form of simplified decision parameters, for example concretized by legal presumptions, or a typification of illegal user behavior.^[353] On the other hand, platforms should generally be allowed to rely on the substantiated factual submissions of the parties to the dispute when examining whether there has been a clear violation of the law, and this, in parallel to the state’s regime on provisional legal protection,^[354] on the basis of a lower degree of conviction (predominant).^[355] In urgent cases, it is also recommended that platform operators establish clear procedural obligations to act and react, which the affected rightsholder can trigger by means of a – non-automated – notification (by independent actors, so-called trusted flaggers^[356] ) of special risk situations.^[357]

1.3.2.3.2 Procedural Obligations under the Notification and Redress Procedure

133. In the context of the notification and redress procedure, the Digital Services Act assumes a fundamentally increased responsibility for clarification on the part of the reporter or rightsholder:^[358] According to this, the illegality of the content must be ‘sufficiently precise and appropriately substantiated.’^[359] In addition to unambiguous information on the electronic location (URL address) where a presumably illegal content is stored,^[360] the reporter basically owes the disclosure of his identity (name and e-mail address).^[361] Furthermore, he must confirm his good faith with regard to the accuracy and completeness of his information.^[362]

1.3.2.3.3 Transparency Obligations

134. The enforcement of state rights and private standards through platforms turns out to be opaque, complex and dynamically open for development in its prerequisites and implementation. Therefore, it structurally disadvantages the addressees of enforcement. This requires the establishment of transparent platform procedures. Corresponding transparency obligations of the platforms manifest themselves, for example, by means of periodic transparency reports^[363] or a right of access for research purposes.^[364] In essence, they are intended to ensure preventive public scrutiny of platforms – whether by public authorities (to protect competition and consumer interests, for example) or by academia, the media and non-governmental organizations. In addition, clear provisions in the general terms and conditions of the online platforms on the sanctions available should make them predictable for the individual platform user. The factually limited perception of general terms and conditions by users is a general problem that has also been diagnosed and controversially discussed in other fields (such as consumer protection law^[365]),^[366] and to which the Digital Services Act has already responded in a regulatory way.^[367] This finding does not, however, fundamentally call into question the usefulness of transparency obligations for platforms.^[368] One reason being the public control function of such obligations, but also the fact that algorithmic – and in this respect dynamic – decision-making (especially through the use of artificial intelligence in the form of machine learning and deep learning) is practically unpredictable for users.^[369]
135. In this light, online platforms should report on the course of the platform procedures themselves. By way of example, information should be provided in advance as to whether automated systems are used,^[370] in particular whether or in which cases algorithmic decision-making^[371] and upstream content monitoring takes place. The obligation to notify should also extend to the type of sanctioned violations (platform’s own standards and/or state law) and to what extent they are actually punished within a certain reporting period.^[372] Platforms should further provide information on the performance of the algorithms used for decision-making, including their (assessment) evaluation premises.^[373] Moreover, whether individual users or user groups are treated differently from the outset, for example with regard to qualitatively graduated access to a platform’s internal complaints procedure. The reporting obligation should also include whether and on what basis a platform has taken measures at the instigation of a public authority,^[374] how the internal complaints procedure is designed^[375] and how it has been used in detail, especially with regard to the basis and number of complaints as well as the type, number and (median) duration of the platform decisions issued in this regard.^[376] Finally, a platform’s decisions on the removal and blocking of access or downgrading of content deemed to be in violation of the law and standards – whether or not such decisions are made – should be stored in a publicly accessible database.^[377]
136. Regarding the AI Act of the European Union,^[378] which will apply from 2 August 2026, its Art 13 obliges transparency and the provision of information. This obligation is not aimed at informing platform users whose legal sphere may be restricted by AI decisions. Rather, the transparency obligation benefits persons who use a high-risk AI system under their own responsibility (so-called deployers^[379]). For them, the operation of a high-risk AI system should be sufficiently transparent so that the outputs of the system can be interpreted and used appropriately.^[380]

1.3.2.3.4 Information and Justification Obligations with Regard to Individual Moderation or Enforcement Decisions

137. In contrast, information and justification obligations of platforms relate to individual moderation and enforcement decisions. They are primarily intended to improve the reviewability of such decisions by way of appeal, and also to counteract hidden sanctions by the platforms such as shadow banning or downranking. The aim of information and justification obligations is thus to ensure effective protection of individual rights, which is regularly done downstream.^[381]
138. As a consequence, enforcement addressees should have a claim against online platforms to disclose the respective, possibly user-individualized basis for decision-making (such as information pertaining to a user’s credit score and the data used to calculate it) – as well as any changes to them.^[382] This includes disclosing whether automated means were used to process the reports and to make decisions.^[383] Above all, platform operators must disclose the function or cause in which they act on a case-by-case basis in individual cases or the reasons for which they act, when enforcing its own standards or third-party rights granted by the state.
139. Furthermore, online platforms must promptly inform affected parties – users and rightsholders^[384] – of enforcement measures taken and provide clear, sufficient and case-specific justification for these measures^[385] in a way that ensures effective defense of rights.^[386] The obligation to provide reasons should apply regardless of whether a platform enforces rights granted by the state to third parties or its own standards.^[387] This obligation should be backed up by the right of data subjects to contest incomplete or otherwise inaccurate information.^[388] For reasons of effective legal protection, platforms should also have to justify to whistleblowers why they have not enforced the rights of third parties as well as general terms and conditions that protect third parties and have therefore not complied with a report on the matter.^[389]

1.3.2.3.5 Access to the Procedure

1.3.2.3.5.1 Access to the Notification and Redress Procedure

140. The requirement of effective legal protection grants, among other things, the right of access to a procedure.^[390] Transferred to platform-based law enforcement, this means: Individuals whose rights are allegedly violated can claim to effectively participate in a notification and redress procedure to be established by online platforms.^[391] This includes, for one the right to easy (‘user-friendly’), but above all uniform access to such procedures.^[392] Accordingly, online platforms should be required to decide on reported suspected infringements objectively and free of arbitrariness on a uniform basis.^[393] In this context, the requirement for a fundamentally rapid response to reports is condensed into an obligation to act immediately if the violation of particularly weighty legal interests (such as the life or physical integrity of a person) is plausibly reported.^[394]

1.3.2.3.5.2 Trusted Whistleblowers (Trusted Flagging)

141. Reports of illegal content by trusted flaggers are treated with priority and without delay.^[395] Although Art 22 of the DSA Regulation only covers content that violates state law, the Digital Services Act does not prevent online platforms from setting up an independent system in parallel to (privileged) report content that violates a platform’s standards.^[396] The lack of regulation in this area is nevertheless to be criticized, especially in light of the particularly sensitive integration of Internet Referral Units (IRU)^[397] into the platforms’ privileged reporting systems.
142. In addition to a form of civil society supervision,^[398] trusted whistleblowers represent an instrument to partially compensate for the structural lack of independence of platforms in the enforcement of third-party rights^[399] and to remedy illegal content more quickly.^[400] According to the DSA regime, the status of trusted whistleblower is granted upon application by so-called coordinators for digital services in the whistleblower’s Member State of domicile. The status must be linked to certain conditions, such as, a particular expertise of the whistleblower and his or her independence from online platforms.^[401] With regard to the independence requirement, institutions are generally more suitable for this purpose than individuals.^[402] Otherwise, it is not necessary to act in a strictly altruistic manner to protect collective interests,^[403] so that the expertise of interest groups of rightsholders can also be used as prioritized whistleblowers. Public institutions such as government agencies^[404] can (and must) also apply for the status of trusted whistleblower under the Digital Services Act model, thus lending (additional) weight to their legal assessments.^[405]
143. Reporting obligations to be published as well as obligations of the platforms to notify notoriously inaccurate or insufficiently substantiated reports ensure the privileged position of whistleblowers and a certain control over their activities.^[406] The legal consequences of privileged reporting on the (sanction) decision to be taken by the platform vary depending on the codification: While the Digital Services Act leaves this question open, the German Copyright Service Providers Act (UrhDaG) confers a kind of (preliminary) presumption of correctness on notifications from so-called trustworthy rightsholders: corresponding declarations by the rightsholders are capable of rebutting the presumption of legally permitted uses established by § 9(2) UrhDaG. Consequently, the service provider is obligated in accordance with § 14(4) UrhDaG to block content that significantly impairs the economic exploitation of a work on a provisional basis, ie until the internal complaint procedure has been completed. In order to increase the performance of trusted flagging in the context of content moderation, it is proposed to supplement the notice systems, which are still designed for individual cases, with notice and stay down mechanisms with regard to comparable and future content.^[407]

1.3.2.3.5.3 Excursus: The Independence of Platforms Vis-À-Vis The State

144. In the enforcement of rights and standards, it is also necessary for online platforms to act – in economic, financial and personnel terms – independently of both reporting persons and their users. This applies particularly to the relationship with the state, in order to prevent biased behavior in the form of privileged treatment of state enforcement interests.
145. However, the requirement for independence from state actors is not so much to be secured by increasing the responsibility of platforms to provide information about the state’s reports and enforcement concerns.^[408] Rather, tighter and at the same time more appropriate limits should be placed on the instrumentalization of platforms in the execution of state interests. Since, it is important to prevent government agencies from acting beyond their legally determined sphere of competence and responsibility by means of private enforcement actors (such as platforms) and thus overstepping the (constitutional) limits of democratic legitimacy and the rule of law – often outside the scope of judicial review. According to the principle of legality (lawfulness of the administration^[409]), the state may only intervene in the legal sphere of the citizen on the ground of a formal law or another legal norm based on this (reservation of the law);^[410] it must use the instruments and procedures provided for by (administrative) law for its (interventional) actions and, in doing so, take into account the fundamental rights of the addressees of the enforcement (as well as competing guarantees in favor of third parties) affected in the individual case.^[411]
146. If follows from this: In case of alleged illegal action by platform users, the state is in principle only permitted to sanction this action with the means and in the forms of (police) law, for example by issuing deletion or blocking orders to the respective provider of hosting services. Beyond this legal framework, the state should only be allowed to enforce its enforcement interests to protect private rights by means of decentralized, platform-based reporting and enforcement systems (if only for the purpose of reporting alleged infringements) if such action is expressly permitted by law.^[412] In that regard, states are called upon to regulate the reporting process for state whistleblowers:^[413] When using private enforcement agents to (provisionally) sanction rights violations, state actors should, in principle, be subject to the same formal and substantive requirements as within an exclusively state action and enforcement space. If they fail to meet these requirements, the notice and takedown mechanism – underpinned by liability law^[414] – should not be allowed to intervene.
147. The situation is different when state authorities are solely concerned with using platforms to prevent the behavior of users or certain content created by them which, although presumed to violate a platform’s own standards, is nevertheless permitted by law.^[415] In this case, the state already exceeds its competence to act with a reference to lawful content directed at the platform, by interfering without a legal basis in areas of freedom that are protected by the law. Such interference should be prohibited by law. If state actors nonetheless make their enforcement concerns known through corresponding notices, platforms would have to document these and publish them – for example, within the framework of transparency reports.^[416] Platforms enforcing their own standards in such cases should also be required to disclose any notifications from the state to the addressee and to justify the enforcement solely on the basis of a violation of the platform’s own standards. In order to take into consideration the right to an effective legal remedy, the internal complaints procedure must be open to the addressee against the enforcement measure without distinction.

1.3.2.3.5.4 Uniform Access to Internal Complaints Procedures

148. Apart from that, all persons affected by measures to enforce rights and a platform’s own standards (users, other rightsholders, whistleblowers) should have equal access to internal complaints procedures. At this stage of the procedure, Art 20(4) of the DSA Regulation, for example, requires procedural equality.^[417] The principle of equality of legal protection borrowed from the states’ codes of (civil) procedure also means that the order and speed of processing must be determined carefully and without discrimination, taking into account the fundamental rights and legitimate interests of the persons concerned.^[418] Exceptions to this principle are justified in the case of particular urgency of the decision for the complainant;^[419] depending on the individual quality of the complaints, the accuracy, scope and processing time of the complaint decision may vary.^[420]
149. It is not compatible with these requirements if users are not given the opportunity to take action against a platform’s decisions by way of an internal complaint within the framework of special procedures (such as Meta’s escalation procedure).^[421] The same applies to the lack of legal protection against the application of unwritten exceptions (so-called allowances)^[422] or against so-called scaling decisions^[423].

1.3.2.3.5.5 Uniform Access to the Nature of Decision Making?

150. To be separated from this problem is the question of whether online platforms are already required at the level of legal or standard enforcement to structure the way decisions are made in the same way for all (groups of) enforcement addressees. Beyond explicit legal regulations, the fact that operators are in principle free to structure their platform according to parameters that follow their own economic interests and therefore also to shape the type of decision-making differently depending on the targeted user (group) speaks against such a requirement for equal treatment (eg, with regard to a human decision-maker).^[424] Irrespective of an ‘increased commitment of platforms to fundamental rights’,^[425] this manifests their entrepreneurial freedom (Art 16 EU EU Charter), including freedom of contract, which is guaranteed by the EU Charter of Fundamental Rights.^[426] This finding is consistent with the common practice of platforms to base initial decisions, ie the enforcement of standards or third-party rights, vis-à-vis users classified as ‘economically less valuable’, on merely algorithmic decision-making.
151. However, one of the limits to the admissibility of this ‘justice by algorithm’ must be seen in the need for effective legal protection. For effective protection of user rights would no longer be guaranteed if algorithmic decision-making were not subject to any human or manual review by the platform.^[427] As also provided for in the EU’s Digital Services Act,^[428] platform operators should be required to provide appropriate safeguards as part of internal review procedures.^[429] This is particularly important in view of the performance limits of algorithmic decision-making systems, which are becoming apparent particularly in the case of a highly context-dependent evaluation of legal issues.^[430] The use of human decision-makers is also likely to strengthen trust in platform-based legal and contract enforcement overall and contribute to its increased acceptance.^[431]

1.3.2.3.6 Proceduralization of Decision Making

1.3.2.3.6.1 Requirements for the Decision Maker

152. To be able to evaluate particularly context-sensitive issues in a normatively correct manner, decisions on platform-based legal and contract enforcement require the deployment of technically qualified personnel with a certain range of language skills.^[432] Beyond these key suitability requirements, platform-based decision-making structures can be designed in various ways: eg, in the form of collegial bodies,^[433] by involving external expertise^[434] and/or through user participation.^[435]

1.3.2.3.6.2 Decision Parameters and Decision Consistency

153. Recently, the decision parameters for (algorithmic) enforcement of third-party rights and platform-specific terms and conditions have become the subject of state regulation. The Digital Services Act of the European Union, for instance, obliges platform operators to act diligently, objectively, non-arbitrarily and proportionately when taking enforcement measures. In particular, legitimate interests, (fundamental) rights and freedoms of all parties involved^[436] and the freedom and pluralism of the media must be taken into account.^[437] In the literature, it has also been demanded that the sanctions imposed by platforms being ‘perceptible’ for the user acting in violation of pertinent rules (deterrence of sanction).
154. In this context, further developments will show to what extent the requirements of objectivity and freedom from arbitrariness (and thus also the fairness characteristics of proportionality and equality) will induce online platforms to gear their enforcement measures more strongly to the ‘decision-making ideal’ of uniformity or consistency and, therefore, less to the (special) interests of individual users or user groups – precisely in the interest of increased predictability and comparability.^[438] Essentially a platform-based consistency and comparability of decisions is based on two premises: First, that comparable enforcement matters are subsumed under uniform abstract criteria for decision-making;^[439] second, that this is done in a uniform, permanent manner with regard to the standard of interpretation as well as the depth of review, without regard to the person – be it in the enforcement of rights set by the state or a platform’s own standards, be it in the internal review of enforcement decisions in the complaints procedure.^[440] A problem of lack of consistency arises, not only when one of these conditions is missing, but equally when platforms apply unwritten exceptions, ie, exceptions not specified in their terms and conditions, to forms of conduct that are in themselves sanction-relevant to specific circumstances without disclosing this to the affected users or whistleblowers. The ‘allowances’, which are exclusively used in Meta’s ‘escalation procedure’, serve as a model for this:^[441] The Facebook Oversight Board has repeatedly criticized the lack of disclosure of their objective, content-related and temporal application requirements as a violation of the principle of legality (sic!).^[442] There is also a risk of inconsistent decision results because Meta links an in-depth, context-related examination of facts by particularly expert decision-makers to specific conditions that are not disclosed to the outside world, such as the fact that government agencies act as reporters.^[443]

1.3.2.3.6.3 ‘Scaling’ of Decisions

155. One instrument for increasing a platform-wide formation of decision standards and decision consistency is so-called scaling. This is understood to mean the (technical) possibility for platforms to apply the concrete standards of interpretation obtained in a ‘model case’ and considered to be of ‘high quality’ to comparable circumstances.^[444] Scaling thus allows online platforms to quickly adapt and standardize their decision-making practice with regard to new decision-making situations that are similar in a large number of cases and comparable in this respect, not least with the aim of providing more effective protection against systematic human rights violations.^[445] If the legal situation is correctly assessed, this will undoubtedly promote effective legal protection.^[446]
156. The disadvantages of such scaling processes are obvious: The more context-sensitive disputes are – as for example in the law of statements – the more likely it is that different cases will be treated equally without any objective reason. Distortion effects are more likely to occur if scaling is not strictly party-related but ‘supra-individual’: This might become relevant if, in the case of defamatory content, a platform were to remove not only those terms or images classified as violating the law or the general terms and conditions that relate to the party in the initiating proceedings, but also, in addition, corresponding content that – in a situational context that is related but not normatively equivalent – relates to other persons. This poses the risk of overenforcement, as well as the opposite effect, if exceptions are scaled excessively (‘scaled allowances’).^[447]
157. As practice shows, scaling can also be based on the decisions of external review bodies. For example, in the case of decisions by the Facebook Oversight Board (FOB) – an external body that also takes into account external standards of interpretation such as human rights when monitoring the platform’s own general terms and conditions – the Meta Group undertakes to examine whether similar situations exist on its platform and whether individual board decisions can be applied to these cases.^[448] In this respect, the Meta Group has a wide scope for estimation.^[449] However, the more individual and context-sensitive the legal disputes are, the smaller the scaling effect of a FOB decision is likely to be.^[450]
158. Scaling issues also arise at the level of law infringement. In the Glawischnig case, for example, the CJEU ruled that hosting providers are obligated to delete content that violates personal rights in parallel cases, provided that the information is identical in wording or – according to the principle – only identical in meaning.^[451] In contrast to the scaling of decisions by the FOB on the interpretation of the platform’s own standards, this case law, which is aimed at effectively protecting personal rights, can be based on the authority of a state court decision. Admittedly, the duty of a platform to remove content that violates state law worldwide and thus extraterritorially, or to block access to it, collides with foreign state sovereignty and consequently with the requirement of consideration under international law in view of nationally (strongly) diverging value systems (especially in the law of expression).^[452] Furthermore, if the scaling refers to decisions made by non-governmental dispute resolution bodies (or even platform courts) on the presumed infringement of legal rights, it is to be feared, detached from questions of extraterritorial extension of effect, that in a large number of cases the interpretation and guiding effect of state law beyond the judiciary’s responsibility for jurisdiction would be modified, and even undermined in extreme cases.
159. In general, the dangers described highlight the need for transparency and downstream effective legal protection:^[453] Complaints bodies must be enabled to review facts on the basis of a scaling decision disclosed to the affected parties in a context- and case-specific manner. The more information-intensive the facts and the more context-sensitive the legal situation, the more reluctant platforms should be to use scaling.^[454] Conversely, the acceptance of scaling decisions is likely to be higher if they are based on external objective review standards and independent review bodies.
160. The platforms’ reluctance to disclose scaling may be due to the fear that disclosure of scaling will lead to a regional differentiation of a platform’s standard terms and conditions, which are often uniform across regions, and that this differentiation will be visible to the outside world. This is precisely the effect that is likely to accompany a corresponding disclosure obligation.

1.3.2.3.7 Excursus: AI Act and Platform Enforcement

161. One question of practical importance is whether the AI Act of the European Union,^[455] which will be applicable from 2 August 2026, will set limits to the AI-based enforcement of standards and third-party rights by online platforms. This is because providers and deployers of AI applications,^[456] which are categorised as high-risk AI systems, are subject to special obligations, including inter alia: Establishing a risk management system (Art 9 AI Act); complying with data and data governance requirements, in particular with regard to training, validation and testing data sets (Art 10 AI Act); furthermore, obligations for technical documentation, record-keeping and transparency (Art 11-13 AI Act) and to ensure human oversight (Art 14 AI Act) as well as an appropriate level of accuracy, robustness and cybersecurity over the entire lifetime of AI systems (Art 15 AI Act). In addition, providers of high-risk AI systems must ensure that their systems fulfil the requirements of the Regulation, Art 16(1)(a) AI Act, in particular that they have a quality management system in place, Art 16(1)(b), Art 17 AI Act.
162. Whether AI systems are to be classified as high-risk AI systems is determined, among other things, by Art 6(2) in conjunction with Annex III AI Act. In the area of ‘administration of justice and democratic processes’, this applies to ‘AI systems intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution’.^[457] The inclusion of such systems in the ADR represents an extension compared to the proposal of the AI Act, which originally only covered their use by judicial authorities.^[458]
163. This raises the question of whether ‘ADR’ within the meaning of the AI Act also refers to algorithmic rights and standards enforcement through online platforms. This question must be answered by means of a of an autonomous interpretation of the law of the Union.^[459] Firstly, it should be noted that the AI-based enforcement of rights and standards by online platforms is not ‘purely ancillary administrative activities’ within the meaning of recital 61 s 5 AI Act, which the Regulation excludes from the group of high-risk AI systems. This is because the examples listed there – such as anonymisation or pseudonymisation of judicial decisions, documents, or data – differ categorically in their legal (internal) effects from a platform enforcement aimed at external effects vis-à-vis users or third parties. Furthermore, although the term ‘ADR’ is mentioned in Annex III no 8 lit a) and in recital 61 AI Act, it is not legally defined here or elsewhere in the Regulation. However, its basic conception under EU law is reflected in the ADR Directive.^[460] Without specifying a particular dispute resolution method,^[461] the ADR Directive defines alternative dispute resolution as ‘a simple, fast and low-cost out-of-court solution to disputes’.^[462] Accordingly, a constitutive feature of out-of-court dispute resolution is the independence or impartiality of the organisation or person appointed to make the decision^[463] – a principle that the AI Act also expressly assumes in its recital 61 s 1 and 4.^[464] In light of this, the term ‘ADR’, also from the perspective of the AI Act, is to be understood as a form of dispute resolution by independent out-of-court decision-makers which – like ADR entities within the meaning of the ADR Directive^[465] – are similar to state courts in their function and basic structure of independent decision-making.^[466] This deliberately excludes forms of dispute resolution that are operated by companies themselves.^[467]
164. As a result, there are important reasons against categorizing platform enforcement as an instrument of alternative dispute resolution. This is due to the fact that platform enforcement is not directly aimed at resolving disputes in the form of a decision^[468] nor does it provide for the use of independent dispute adjudicators. Unlike court proceedings, which typically aim to (provisionally) terminate legal disputes, platform enforcement involves its own concern, namely to unilaterally (if only provisionally) enforce standards or third-party rights on the basis of an independent assessment of the legal and factual situation. In that regard, online platforms are not able to guarantee the independence of the decision-maker required by the above regulations, neither when enforcing their own standards nor when enforcing the rights of third parties.^[469] They can therefore be ruled out as institutions for the ‘administration of justice’ according to Annex III no 8 lit a) of the AI Act.
165. A different assessment could only be reached if the ADR concept of the AI Act were to be interpreted broadly, ie teleologically and on the basis of an effects-based approach, in contrast to the understanding of ADR under EU law. However, such an assessment cannot be derived from recital 61 s 3 of the AI Act, which only classifies AI systems used in ADR as high-risk ‘when the outcomes of the alternative dispute resolution proceedings produce legal effects for the parties’.^[470] The fact that this can only refer to the binding effect of ADR decisions, but not to any impairment of existing rights and legal interests of platform users or third parties, results on the one hand from the functional-structural proximity of ADR proceedings to court proceedings – a proximity that is also reflected in the reference in recital 61 s 3 AI Act to the legal effect that the results of the alternative dispute resolution produce for the parties to the proceedings. A further point of reference arises from Art 86(1) AI Act: The provision confers a right to explanation of individual decision-making if this is based on a high-risk AI system and the decision ‘produces legal effects or similarly significantly affects that person in a way that they consider having an adverse impact on their health, safety or fundamental rights’. Accordingly, the provision presupposes the use of high-risk AI systems within the meaning of Art 6, Annex III AI Act, for example in the course of an ADR decision with binding effect on the parties. This effect must differ from the effects of a decision referred to in Art 86(1) AI Act – ie the legal or factual impairments resulting from it – which would otherwise have no independent normative significance.^[471] In other words, a right to an explanation under Art 86(1) of the AI Act does not already follow from the binding effect of state or extrajudicial decisions, but solely from the additional circumstance that such a decision interferes legally or factually with existing rights or legal interests of persons in individual cases.
166. To classify platform enforcement as a suitable field of application for high-risk AI systems within the meaning of Annex III no 8 lit a), a purely teleological interpretation of the provision would have to be adopted, by considering the principle of effet utile under EU law. The starting point for such an approach is that the AI Act is designed as a risk-based regulatory system for AI applications.^[472] Accordingly, the riskier the operation of the AI, the more comprehensive the obligations in connection with the operation of the AI.^[473] This regulatory concept, which is tailored to the intensity and scope of the risks, is manifested in recital 61 s 1 AI Act, among other things. Here, the Union legislator always advocates classifying AI systems as high-risk if they are likely to have a ‘significant impact’^[474] on individual freedoms and the fundamental rights protected by the Charter – such as the right to freedom of expression or the right to an effective remedy^[475]. The severity of the harm and its likelihood of occurrence must be taken into account (recitals 48 s 2, 52 s 1 AI Act). The previous explanations have shown^[476] that the potential for errors and the type, scope and probability of imminent risks and damage when AI systems are used in platform enforcement are typically no different or even less pronounced than in the fields of application addressed in Annex III no 8 lit a). In view of the procedural structures proposed for platform enforcement,^[477] it is not apparent that the obligations of providers and operators of high-risk AI systems set out in Art 9-17 AI Act would impose a fundamentally unreasonable hardship on online platforms.

1.3.2.4 Necessary Dovetailing of Internal Complaint Procedures with Out-Of-Court Independent Dispute Resolution

1.3.2.4.1 Structural Bias of Online Platforms

167. Online platforms are sometimes described as ‘objective’ dispute decision-makers similar to a judge or even like a judge.^[478] Only recently, such a role description is echoed in the Digital Services Act, when online platforms are required to process complaints submitted by users as part of a newly established, internal legal protection procedure in an ‘objective manner’.^[479]
168. However, this role model loses sight of the fact that platforms typically act in a liability-based and interest-driven manner and are thus structurally biased.^[480] This becomes obvious when platforms do not enforce third-party rights granted by the state, but rather – often automatically – their own community standards, ie general terms and conditions set by the platform itself and – as in the case of Facebook^[481] – applicable worldwide.^[482] In this quantitatively far predominant enforcement situation,^[483] the platform no longer acts as a mediator between third-party interests, but as an enforcer or – in complaint proceedings – as a judge in its own cause.^[484]
169. By the same token, it is hardly surprising that platforms have generally shown little inclination so far to remedy complaints against the enforcement of self-imposed standards or – in the absence of liability-reinforcing submission obligations^[485] – to submit them to an independent extrajudicial body.^[486] In Germany, for example, Facebook made use of a corresponding right of referral in only 29 of a total of 123.195 complaints under the Network Enforcement Act (NetzDG) in the 3rd and 4th quarters of 2022, ie in just around two out of 10.000 cases.^[487] The figures for other platforms are similar – or even more drastic.^[488]
170. In this context, effective incentives to increase the platforms’ willingness to submit documents can also be provided by time limits that are subject to liability or even fines.^[489]

1.3.2.4.2 Out-Of-Court Dispute Resolution

171. In that regard, internal platform-based complaints procedures should be supplemented by external, out-of-court dispute resolution and these two protection instruments be more closely interlinked.^[490] Only when internal and external complaints procedures are intertwined can an effective legal remedy be guaranteed (see recital 52 p. 2 of the DSA-Regulation). A right to appeal is intended to enable affected users and whistleblowers to address a certified independent dispute resolution body directly (and as an alternative to appealing to state courts^[491]) after an unsuccessful internal complaint against a decision by the platform about content that is allegedly contrary to law or the terms and conditions,^[492] even against the will of the platform.^[493] The decision of the dispute resolution body – which is not to be automated – should in turn be binding for the platform,^[494] at least until a possible state court decision.^[495] It should also be stored anonymously and in an easily accessible manner and published online.^[496]
172. If, according to the controversial and criticised concept of the DSA, decisions of certified out-of-court dispute settlement bodies do not have any binding effect on the parties, ^[497] AI systems that these bodies may use to assist in researching and interpreting facts and the law and in applying the law to a concrete set of facts, are not to be categorized as so-called high-risk AI systems within the meaning of the AI Act of the European Union.^[498] As a consequence, the providers and operators of such high-risk AI systems are not subject to the special obligations under Art 9-17 AI Act.^[499]
173. In addition, out-of-court dispute resolution (eg, in the context of Art 21 DSA Regulation) opens up an incidence check of the platform general terms and conditions – and the more fluctuating moderation guidelines^[500] that concretize them – with regard to their conformity with the procedural and contractual requirements^[501] of pertinent EU regulations (Digital Services Act; P2B Regulation).^[502] In view of the complex and numerous requirements of EU law, from the perspective of platform operators the law on general terms and conditions is turning into a freedom-restricting vehicle for inclusion in private platform usage contracts. Platform general terms and conditions that have been overhauled by state law are taking on a regulatory function. In this way, however, the state instrumentalizes private regulatory potential even in areas that – for reasons of protected spheres of freedom of expression – are typically removed from the state’s legislative competence.^[503]

1.3.2.4.3 The Establishment of ‘Platform Courts’

174. Even before the Facebook Oversight Board was set up, there were suggestions that a platform-based enforcement of rights and, above all, of a company’s own standards and terms and conditions should be flanked by external expertise, namely in the form of non-governmental, court-like dispute resolution bodies.^[504] Corresponding ‘platform courts’ will not only have the task of settling individual disputes (usually of particular relevance) with particular proximity to the platform’s subject matters and continuously harmonizing decision-making and interpretation of the community standards of a platform.^[505] Rather, they are also assigned a regulatory function: In implementing what is essentially a ‘procedural concept of (self-)regulation’, platform courts shape a private process of norm-building that is (partly) beyond a state’s control of general terms and conditions and is communicatively linked back to the public.^[506] In this process, the dimensions of mass and time that are typical of platforms (for example, with regard to ephemeral communication and evaluation processes) must be managed in a regulatory manner.^[507] In the interest of a (mostly heterogeneous) mass of platform users, the aim is twofold: on the one hand, to ensure more effective protection of individual rights in a multitude of (parallel) cases and, on the other, to enable a continuously adapted, by its nature mostly rudimentary-experimental rule formation through and within the framework of procedural structures.
175. The conceptual approaches advocated in this context are linked in different ways: either effect-related to a norm-building process through (limited) precedent;^[508] or dogmatically through its foundation on ‘treaty networks’;^[509] or institutionally, for example in the form of an independent, cross-platform ‘court of justice’ for the formation of a (sectorally differentiated) substantive ‘platform common law’.^[510] Parallels can also be drawn functionally to the further development of law by the appellate courts or, more concretely, to the functioning of the WTO Appellate Body.^[511] In the absence of statutory regulation, such projects presuppose that platforms voluntarily submit to private platform jurisdiction – and its indirect regulatory effects – at least within certain business areas. Irrespective of how closely the relevant dispute resolution bodies are organizationally linked to a platform or how their internal structure is differentiated by various normative functionaries, interpretation (dispute-related review), setting and enforcement of private standards ultimately amalgamate into ‘one power’.

2 Artificial Intelligence in Mediation and Conciliation

176. This chapter takes a closer look at the use of AI in mediation and conciliation. There is no uniform understanding of the exact terminology used in the two dispute resolution systems. For mediation, the definition of the EU Mediation Directive^[512] will be used in the following. It defines mediation^[513] in Art 3 lit a) as a ‘structured process [...] whereby two or more parties to a dispute attempt by themselves, on a voluntary basis, to reach an agreement on the settlement of their dispute with the assistance of a mediator. This process may be initiated by the parties or suggested or ordered by a court or prescribed by the law of a Member State.’
177. In contrast, the conciliator, as a neutral intermediary, plays a more active role in the conciliation process by submitting – albeit non-binding – proposals for decisions.^[514] Conciliation is therefore a third-party decision procedure.^[515]
178. Then again both procedures pursue the goal of an amicable, in principle non-binding dispute resolution,^[516] under the intermediary intervention of a neutral third party, regularly detached from substantive and normative requirements.^[517]

2.1 Fields of Application of Artificial Intelligence

2.1.1 Fields of Application According to Process Stages

2.1.1.1 Procurement of (Legal) Information by Artificial Intelligence

179. Legal chatbots are already able to provide low-threshold information for those seeking legal advice in the run-up to mediation and conciliation proceedings – for example about the course of proceedings or the costs involved.^[518] Such as answering simple legal questions,^[519] whereby, depending on the jurisdiction, professional restrictions must be observed.^[520] However, the use of such large language models^[521] carries the risk of providing incorrect information (so-called ‘hallucinations’), which are not easily recognizable as such to a layperson.^[522]

2.1.1.2 Automated Document Analysis, Fact-Finding and Case Management

180. Similar to arbitration proceedings, the use of automated document analysis also promises to reduce the workload for the parties involved in mediation and conciliation. One possibility is the automated structuring of party submissions, whereby the facts of the case are recorded in advance using input masks.^[523] Further, AI-based automation can be used to classify the status of disputes, for instance in the categories of child maintenance or contact issues. The system can then assign the dispute status prepared this way to a human decision-maker, possibly taking into account the urgency of a decision (‘triage’), which is also determined by AI.^[524] In the form of smart assistants or smart scheduling, there is also potential for automation in the further conduct of proceedings. The use of AI-supported speech recognition and translation is also conceivable,^[525] as are, in principle, potential applications for an automated selection of mediators or conciliators.^[526]

2.1.1.3 Ai-Tools for Preparing Settlement or Decision Proposals (Predictive Tools)

181. The use of predictive tools in mediation and conciliation bears potential. In Singapore, for example, there is a procedural simulator intended to predict the outcome of legal – court^[527] – disputes for the parties and thus increase their willingness to settle.^[528] The use of such technology is being discussed in proceedings before consumer conciliation boards, too. Its purpose is to encourage consumers to actually submit their complaint.^[529] Human conciliators are also beneficiaries, as AI systems create draft texts or drafts of conciliation proposals (document drafting and pre-drafting).^[530] Parties also receive support when drafting an agreement in mediation proceedings.^[531] Finally, the proceedings^[532] – including the decision-making process and in case of conciliation – could possibly be conducted entirely by machine, although the technical feasibility of this is not yet apparent at present.^[533]

2.1.2 Sectoral Fields of Application of AI in Mediation and Arbitration

182. Platform-based conciliation serves as inspiration for an advanced automation of out-of-court dispute resolution. Large intermediaries not only mediate the exchange of goods, but also take on the role of a third-party adjudicator with respect to the relationship between their customers – eg, between sellers and buyers on a platform (‘third-party-adjudication’).^[534] By way of example, the chargeback process of the payment service provider PayPal^[535] can be fully automated with the help of the third-party provider chargeflow: The system automatically recognizes potential chargeback cases, calculates the chances of success, automatically drafts the pleading with corresponding evidence and submits it to PayPal. Further, the appeal against PayPal’s initial decision will be fully automated (so-called charge appeal).^[536]
183. In the resolution process of the eBay Resolution Center, an algorithm mediates between the parties to the dispute; 90% of cases are resolved purely automatically in this way.^[537] As the entire transaction is processed via eBay, the Resolution Center has direct insight into the facts of the case.^[538]
184. In contrast, the degree of automation of state-recognized^[539] and purely private conciliation bodies as well as ODR providers^[540] is limited to electronic communication.^[541] This applies in particular to the area of family mediation.^[542]
185. Despite all these developments, as far as can be seen, AI is currently only used in the procedure simulator in Singapore and in a comparable tool in British Columbia (‘Solution Explorer’).^[543] The latter involves the use of a simple, rule-based AI (so-called expert system).^[544]

2.2 Use and Dangers of Using Artificial Intelligence

2.2.1 Benefit

186. The (hoped-for) benefits of using AI are obvious in the area of mediation and conciliation as well. On the one hand, procedural efficiency gains are expected, expressed, for instance, in faster processing of applications and similar disputes (eg, before consumer conciliation boards).^[545] This is usually accompanied by a lower cost burden compared to manual processing by humans.^[546] On the other hand, it remains to be seen whether automated or AI-based instruments will equally facilitate access to out-of-court dispute resolution (eg, in consumer matters) and be able to overcome the rational disinterest of consumers^[547] even in countries where the risks of legal action are better covered by legal aid and legal expenses insurance.^[548] To what extent and in which areas automated online dispute resolution systems are desirable in terms of legal policy is beyond a uniform and – in view of constant further developments – conclusive assessment at the present time. In addition, the attractiveness and actual use of such forms of dispute resolution stands and falls with the relief of the state courts.^[549]

2.2.2 Dangers

2.2.2.1 The Black Box Problem and Discrimination

187. Similar to platform-based law enforcement and arbitration proceedings, the use of AI harbors the risk of opaque decision-making bases, particularly in the area of dispute resolution (black box problem). Even assuming that modern systems are capable of processing new, yet unknown data after the learning phase, such modus operandi inevitably leads to an output becoming independent from the input, so that the individual steps leading to a decision can no longer be reconstructed. This essentially concerns the sources and method of decision-making, ie on the one hand the way in which the multitude of stored information is linked and on the other, and above all, whether the system has merely reproduced a certain normative program or ‘added’ something to it.^[550]
188. By the same token, it is hardly possible to explain and justify fully automated decisions.^[551] Along with this goes a possible loss of legitimacy of AI-based decision-making proposals or suggestions, which may affect the acceptance of such procedures and their decision-makers as a whole.^[552] If AI is not used solely for informational purposes in the run-up to mediation or conciliation^[553] or with the aim of automated document analysis or fact-finding, but (at the same time) with the intention of supporting decision-making in mediative compromise finding or the preparation of conciliation proposals,^[554] there is a risk of discriminating against parties to proceedings (in individual cases) and undermining independent and impartial decision-making with regard to certain facts or groups of persons concerned.

2.2.2.2 ‘Petrification’ of Decision-Making through Strict Reference to the Past

189. A key point of criticism of the use of AI in mediation and conciliation, as in arbitration proceedings, is the finding that, for technical reasons, AI is not capable of making cognitive and (spontaneous) value-based decisions^[555] or showing empathy – as a cornerstone of interpersonal communication. AI also lacks self-confidence and the ability to explain its own algorithms.^[556] The reason for this is obvious: Given that training data is only able to represent the past,^[557] AI knows nothing about the current physical world and therefore about present life in general. Therefore, AI lacks a direct and immediate view of this; judgment, intuition, (legal) feeling, aequitas and tact are necessarily alien to it.^[558] This disadvantage always becomes apparent when improbable events occur or the existing body of norms mapped in the algorithm is not sufficient and must therefore be supplemented by interpretation or further development of the law.^[559]
190. Rather, by referring to past case material, AI makes ‘conservative’ decisions and is therefore not technically capable of developing the law, especially in the case of legal situations strongly depending on values and individual circumstances.^[560] Past-oriented decision-making is particularly problematic in the event of changes to the law. The reason is that decisions made on the basis of an old legal situation become incorrect as training data and therefore can no longer be used.^[561] The ability of artificial intelligence to predict and adapt cannot compensate for the underlying technical problem,^[562] although the low publication density in the area of conciliation reduces the severity of the problem.^[563]

2.2.2.3 ‘Legal Remoteness’ Decision-Making Standards and the Accumulation of Private Decision-Making Power

191. In light of this, trust and willingness to (fully) rely on the results of a machine in matters of dispute resolution are generally (still) not very pronounced.^[564] However, this limits the use of AI from the outset to disputes with more similar conflicts of interest that are easier to map technically in terms of the underlying evaluations.^[565] In addition, especially in the area of platform-based conciliation, decisions are often not made (strictly) on the basis of (and to enforce) substantive law:^[566] Normative standards^[567] are replaced by a company’s own standards and general terms and conditions and thus rules being – with a view to the economic interests of the respective user – more focused on fulfilling customer satisfaction than on a finely balanced resolution of conflicts of interest through legal principles or on the execution of legally recognized needs.^[568] As a prominent example of this serves the platform-based dispute resolution of the payment service provider PayPal.^[569]
192. Indeed, the findings of a certain ‘legal remoteness’,^[570] the simplicity of the applicable rules and regulations^[571] as well as a ‘privatization of civil law’^[572] (and its procedural implementation) are conceptually inherent to the dispute resolution forms of conciliation and – in particular – mediation and have therefore been known for a long time.^[573] However, the increased use of AI is accompanied by a shift in emphasis: away from consensual, individualized, case-by-case dispute resolution^[574] towards a more standardized, rule-based, scalable decision-making process in the context of a large number of comparable conflict situations. This, in turn, is associated with the risk of a ‘fossilization’ of dispute resolution, which pushes back the fundamentally necessary possibility of (spontaneous) legal development.^[575]
193. Last but not least, the progressive absorption of proceedings by intermediaries acting as third-party adjudicators (PayPal, Amazon Resolution Center, etc) is leading to an accumulation of private decision-making power in an imperfectly regulated area of law.^[576] All of these developments threaten to weaken the acceptance of mediation and conciliation as a whole.

2.3 Procedural Answers

194. Setting a regulatory framework for the beneficial and risky use of AI is a task of the law that arises in the area of mediation and dispute resolution, too. In terms of implementation, it is (largely) similar to the approaches outlined in the platform-based dispute resolution and arbitration.
195. For one thing, this is linked to the need to create transparency with regard to the use of AI.^[577] The Artificial Intelligence Act recently adopted by the European Union^[578] sets out transparency rules for certain AI systems that are particularly susceptible to manipulation.^[579] These include, for example, AI systems intended for interaction with natural persons, such as chatbots in particular.^[580] The AI Act also standardizes prohibited practices, which may include nudging, dark patterns and behavioral microtargeting.^[581] In this regard, certain obligations are imposed on providers of high-risk systems,^[582] such as the creation of a quality management system or technical documentation for the high-risk AI system.^[583] According to the text of the regulation as amended by the EU Parliament, the information obligations extend to the legal protection afforded by procedures when using AI systems.^[584]
196. For another thing, the establishment of ‘hybrid decision-making’ is being discussed: AI-based proposals for decisions – such as an out-of-court dispute resolution body in accordance with the DSA^[585] or the German UrhDaG^[586] – should never have binding effect without the consent of the parties to the proceedings. If the parties refuse to give their consent, it would be up to the individuals to make a binding decision on the subject matter of the dispute.^[587] In the best-case scenario, this model promises to combine the advantages of human decision-making (especially in disputes with a strong focus on normative values and individual circumstances) with the (potential) cost and resource-related efficiency of automated decision-making systems – a pledge for efficiency that has to prove itself in particular when dealing with a large number of similar disputes (scaling of decisions).
197. Regarding the need for procedural minimum standards in the area of mediation and conciliation, please refer to the comments on platform-based dispute resolution.^[588]
198. Both conciliation proceedings, in which a conciliator submits non-binding proposals for decisions to the parties, and mediation, which leaves the parties to resolve the conflict on their own responsibility, do not fulfil the definition of ADR proceedings that AI Act of the European Union establishes^[589] – due to the lack of binding effect of their decisions on the parties. Consequently, AI systems that a conciliator or mediator may use to assist in researching and interpreting facts and the law and in applying the law to a concrete set of facts are not to be classified as so-called high-risk AI systems within the meaning of the EU AI Act. Consequently, the providers and operators of such high-risk AI systems are not subject to the special obligations under Art 9-17 AI Act.^[590]

3 Arbitration and Artificial Intelligence

3.1 Fields of Application of Artificial Intelligence in Arbitration Proceedings

199. Although the use of artificial intelligence in arbitration proceedings is still limited, the technology has been gaining importance in recent years.^[591] Mainly because such use promises an increase in procedural efficiency, in particular a reduction in costs.^[592] The main current fields of application of AI are the following:

3.1.1 Document Analysis and Case Management

3.1.1.1 Automated Document Analysis

200. Reflecting a general trend, AI is currently predominantly used in the field of automated document analysis.^[593] The focus lies on technologies for data evaluation and structuring.^[594] The fields of application are diverse: In addition to intelligent document searches (so-called smart searches)^[595] and the automated creation of a process history,^[596] this includes the structuring of party submissions, evidence and presentations of evidence^[597] as well as automated e-discovery.^[598] Use is also made of forecasting tools for the automated collection and analysis of market data (eg, in the energy sector) being relevant to the outcome of proceedings.^[599]
201. Speech recognition and translation are further fields of application. With the help of AI tools, transcripts of verbal negotiations can be created automatically and in real time^[600] or translations be produced.^[601]
202. In regard, natural language processing (NLP),^[602] optical character recognition (OCR)^[603] or predictive coding^[604] act as technical tools, although the use of AI is by no means mandatory. In individual cases, less technically complex software is also suitable, for example for automated keyword searches.^[605]

3.1.1.2 Automated Document Creation and Advanced Analysis Tools

203. As in the legal services market,^[606] AI-driven research and analysis of case law may also take place in arbitration proceedings.^[607] For example, tools are already being used to create summaries of arbitral awards and court judgments based on a database of international law and arbitration law documents.^[608] Another field of application is the automated creation of drafts, such as pleadings (pre-drafting).^[609] Similarly, artificial intelligence can be used to review documents in order to avoid legal and factual errors.^[610]

3.1.1.3 (Other) AI-Supported Case Management

204. It is also possible to outsource parts of the organization of the procedure to artificial intelligence. For instance, AI systems are capable of checking deadlines^[611] or schedule meetings and hearings automatically (smart scheduling).^[612] The same applies to checks on whether an arbitration award meets certain formal requirements, such as the signature of all arbitrators.^[613] In addition, the use of so-called smart personal assistants is conceivable, which are tailored to the respective circumstances of the individual arbitration proceedings.^[614]

3.1.1.4 Legal Admissibility

205. Asking about the admissibility of such use of AI,^[615] the supreme personality of the arbitrator’s mandate^[616] must be observed when making use of advanced software. As a consequence, the outsourcing of an arbitrator’s core tasks without the consent of the parties proves to be inadmissible.^[617] A problem arises in particular for auxiliary activities such as the summarization of factual sequences and expert opinions, the review of other documents or the examination of evidence, which arises in a comparable way when deploying tribunal secretaries in arbitration proceedings.^[618] In order to avoid an arbitral award being set aside, it is advisable to contractually regulate the involvement of AI in parallel to the involvement of tribunal secretaries.^[619]

3.1.2 Selection of the Referee Using AI

206. Another auxiliary function of AI is the selection of referees. This is conceivable on the basis of an analysis of specialist databases^[620] or a broader internet search,^[621] combined with the creation of a ranking list of the candidates found. In that regard, the AI is technically able to consider the parameters specified by the parties and, if necessary, considering conflicting criteria, eg, by implementing an expert system.^[622]
207. However, implementation problems arise in view of the extremely incomplete data basis.^[623] Although there are some databases specializing in the evaluation of arbitrators and arbitration institutions,^[624] their analytical capabilities remain deficient as only providing an incomplete and generally unreliable basis for suitable use as training data.

3.1.3 Technologies for Risk Assessment and Prediction of Process Outcomes (Predictive Analytics)

208. Another area of application for AI is risk assessment aiming at a prediction of the outcome of proceedings^[625] or even the behavior of individual arbitrators^[626] (predictive analytics). This puts the parties in a position to assess their chances of success (supposedly more reliably), or at least use the AI prediction as a basis for settlement negotiations.^[627] This technology shows to be of great interest for the rapidly growing market of litigation financing in arbitration proceedings, hoping for a statistically improved prediction of the outcome of proceedings and litigation risks.^[628]
209. The prediction process is of a statistical nature, ie, not a cognitive process involving the application and analysis of relevant legal rules.^[629] Therefore, the decisions are obtained inductively from a large amount of historical data on decisions and cases. The AI system searches for patterns or correlations relevant to the outcome of the decision. Suitable input data is essentially metadata (metadata analysis)^[630] and factual data^[631] (factual data analysis),^[632] being technically implemented by using the random forest method^[633] or natural language processing.^[634]
210. From a legal perspective, the use of such predictive analytics – still being unregulated in arbitration proceedings^[635] – is not uncontroversial. Nevertheless, the use of such tools raises concerns regarding the independence and impartiality of arbitrators, especially when an arbitrator is appointed on the basis of decision predictions.^[636] This raises the intricate question of whether any duty of investigation on the part of the arbitrator – resulting from an obligation of disclosure^[637] – with regard to conflicts of interest^[638] also extends to an investigation into the use of such predictive tools by the nominating party or a third-party financier.^[639] It is equally questionable whether an arbitrator can be challenged because the prognosis shows with overwhelming probability that the case can only be decided by him with a certain procedural outcome. In line with the ‘judge scoring’ now prohibited by law in France,^[640] there are some calls for a complete ban on predictive analytics in arbitration proceedings.^[641]

3.1.4 Fully Automated Process Management and Decision-Making (‘Robo Arbitrator’)

211. Fully automated arbitration procedures in which decision-making is carried out exclusively by AI are still far beyond the current state of the Art Nevertheless, a brief analysis is already worthwhile at this stage, not least due to the rapid pace of technological progress.^[642] Fully automated decision-making raises both technical and legal concerns:

3.1.4.1 Computability of Law?

212. Given that decision-making by AI is purely based on pattern recognition and statistical methods, rather than of a legal methodological subsumption act,^[643] AI is not capable of examining and deciding complex legal situations – which are often the subject of arbitration proceedings.^[644] As a result, a purely AI-based decision-making process would only be suitable for disputes that are uniform in nature and based on easily verifiable facts – often (but not necessarily) with a low value in dispute (small claims). Examples include flight delays^[645] or consumer arbitration, being widespread in the USA.
213. The already established lack of AI systems to make human assessments^[646] and show emotions and empathy^[647] is even more pronounced in arbitration proceedings, where there is a much greater need for individual case and fairness decisions.^[648] Albeit existing AI being capable of recognizing emotions based on facial expressions and gestures, this technology is still at a rudimentary level. In addition, these parameters would have to be included in the machine decision-making process, thus posing considerable technical difficulties.^[649]
214. In this light, the replacement of referees by AI is not even remotely foreseeable at the present time.

3.1.4.2 The problem of obtaining training data

215. Another technical barrier is the acquisition of training data. Due to the low publication density of arbitration awards, there is little case material available.^[650] Indeed, publicly accessible data sets already exist, including, inter alia, the decisions of the International Center for Settlement of Investment Disputes (ICSID), the Society of Maritime Lawyers (SMA) and the Court of Arbitration for Sport (CAS).^[651] By the same token, organizations have developed databases containing decisions and other data sets (such as interviews with the parties to the arbitration proceedings).^[652] Regardless of whether a trend towards the increasing publication of arbitration awards can actually be observed,^[653] existing data is only suitable for AI-supported analysis to a limited extent due to its incompleteness, eg, with respect to a lack of (exhaustive) reasons and the names of the parties involved, including the arbitrator.^[654]

3.1.4.3 Legal assessment of autonomous decision-making by AI

216. A fully AI-driven decision-making process also raises considerable objections from a legal perspective. First of all, there is no question that outsourcing decision-making authority to AI systems as part of an arbitrator’s personal mandate would require consent by party agreement.^[655] However, even beyond this consent requirement, the question of legal conformity of extensively automated arbitral awards arises. Both international arbitration law and the vast majority of national legal systems assume the requirement that only natural persons or persons with legal personality are suitable as arbitrators. Furthermore, concerns are raised regarding fundamental procedural guarantees – namely procedural equality of arms, the granting of a fair hearing as well as the independence of arbitrators and, not least, data protection.

3.1.4.3.1 International arbitration law: UNCITRAL Model Law and New York Convention

217. While the UNCITRAL Model Law (arguably) assumes that the arbitrator must be a natural person,^[656] there is disagreement about the admissibility of fully automated decision-making systems under the New York Convention on the Recognition and Enforcement of Foreign Arbitral Awards. The wording of the Convention (Art I no 2 NYC) only refers to ‘arbitrators’ as such.^[657] It has occasionally been deduced from this that an arbitrator does not necessarily have to be a natural person, thereby alluding to the purpose of the Convention, ie the cross-border recognition and enforcement of arbitral awards.^[658] With reference to Art IV no 1 lit a) NYC), however, this view is rejected.^[659]

3.1.4.3.2 National arbitration law

218. Some legal systems expressly ask for a natural person as arbitrator: for example, French,^[660] Dutch,^[661] Spanish^[662] and Turkish^[663] arbitration law, as well as Peru, Brazil and Ecuador.^[664] Based on the personal mandate of arbitrators, this also corresponds to the prevailing view in German literature.^[665] Correspondingly, some arbitration rules of the major arbitration institutions assume that a natural person always conducts and decides the proceedings.^[666] Other legal systems explicitly require the legal capacity^[667] of arbitrators, including Sweden,^[668] Italy^[669] and England^[670].

3.1.4.3.3 Formal and data protection hurdles

219. For recognition, Art IV no 1 lit a) NYC requires a certified original or certified copy of the arbitral award. Whether electronic documents are sufficient in this respect is governed by national law.^[671] In this light, there is legal uncertainty as to whether electronic arbitral awards may be subject to recognition at all.^[672] German arbitration law, for example, requires arbitral awards to be signed.^[673] Whether this written form requirement can be satisfied electronically and therefore also by a ‘robo-arbitrator’ is certainly doubtful, given the different regulatory functions of the form requirement.^[674]
220. The requirements of the General Data Protection Regulation (GDPR) also apply to the use of AI in arbitration proceedings.^[675] According to Art 22(1) GDPR, fully automated decisions are prohibited in arbitration proceedings unless the data subject has given their explicit consent.^[676] From a data protection perspective, it would therefore make sense to include the use of fully automated decision-making systems in a party agreement.

3.1.4.3.4 Outlook: Blockchain arbitration

221. An alternative enforcement regime^[677] is blockchain arbitration.^[678] Such decentralized dispute resolution procedures^[679] are characterized by the fact that the contracting parties register the contract in dispute on a blockchain-based platform and deposit a security amount there, which corresponds to the amount in dispute:^[680] The dispute is decided according to the majority principle by a collective of human jurors selected by algorithm, but without them communicating with each other or with the contracting parties. Subsequently, only the jurors who decide in accordance with the majority principle receive remuneration and the decision is implemented immediately through the automated distribution of the deposited contributions.
222. Such dispute resolution mechanisms are often referred to as ‘arbitration’. However, it is very doubtful whether this classification is correct from the perspective of state law and whether it could therefore have binding and blocking effects on state court proceedings. One argument against this is that the current procedures probably do not meet the minimum legal requirements in several aspects.^[681] Therefore, such proceedings may not replace but rather complement arbitration proceedings.^[682]

3.2 Risks Associated with the Use of Artificial Intelligence in Arbitration Proceedings

3.2.1 Unconscious Bias and Risk of Discrimination

223. Using AI merely in a supportive way harbors the risk of unconsciously influencing human arbitrators. Such a scenario is conceivable in the form of a cognitive bias to follow the results of automated reviews (‘anchor-effect’).^[683] In addition, there is a risk of algorithmic discrimination in arbitration proceedings, too, eg, on the basis of gender, ethnicity or age.^[684] The technical cause of this can be the training data, actions of the self-learning system in its environment, but also deliberate manipulation.^[685] The frequently criticized investor-friendliness of arbitration tribunals may serve as a concrete example in the area of investment arbitration. Such a bias in arbitral awards threaten to continue in their future use as training data in the AI model.^[686] Both risks, anchor effects as wells as discrimination, structurally impair the independence of human arbitrators.^[687]

3.2.2 The Black Box Problem

224. A lack of transparency and traceability of automated decision-making become no less virulent in arbitration proceedings. As a matter of principle, the operation and results of AI systems are not comprehensible to users, nor (usually) to developers of the system.^[688] However, if the black box problem prevents sufficient explanation and justification of decisions,^[689] the acceptance of arbitral awards dwindles,^[690] possibly the legitimacy of the decision-making body as such,^[691] but also the behavior-guiding effect of decisions on the parties.^[692] The relevance of a statement of reasons for decisions follows from the fact that parties are, statistically speaking, rarely willing to dispense a statement of reasons or to allow an arbitrator to decide ex aequo et bono^[693] Furthermore, due to the lack of normative reasoning, AI-based decisions are not suitable for use as a reference by other (arbitration) courts.^[694] In contrast to decisions by state courts, however, this problem is likely to arise less frequently.

3.2.3 Danger of ‘Petrification’

225. As artificial intelligence makes decisions on the basis of training data from the past,^[695] the phenomenon of conservative decision-making closed to a development of the law also occurs in arbitration proceedings. This procedural area is by no means alien to a development of the law,^[696] even though the problem described is much less pronounced than in state court proceedings due to the low publication density of arbitration awards.

3.2.4 Failure to Grant the Right to Be Heard

226. It is true that complete automation of proceedings as such does not constitute a violation of the right to be heard. Yet, such a violation must be assumed if significant party submissions cannot be considered in the arbitral decision due to an error in the AI system.^[697]

3.3 Procedural Answers

227. Notwithstanding all technical limitations, AI-based arbitration awards run – at least in some jurisdictions – the risk of being set aside or unenforceable.^[698] In view of the risks associated with the use of AI in arbitration proceedings, regulatory requirements become necessary with the aim of ensuring legal certainty and protect the legal interests of both the parties and the general public. In this context, the principles of transparency and fairness of proceedings as well as the protection of trust in the accuracy of AI systems and the integrity of arbitration proceedings as such should be mentioned in particular, eg, as concerns an AI-based reproduction or summary of the facts, the legal situation or the evidence in a specific proceeding. In that regard, the draft of the Guidelines on the Use of Artificial Intelligence in Arbitration,^[699] recently published by the Silicon Valley Arbitration & Mediation Center, provides useful indications on the design of arbitration proceedings using AI.^[700] In this light, the following regulatory approaches, which are by no means exhaustive, should be considered:
228. In view of an arbitrator’s personal responsibility for the decision-making process, any transfer of his personal mandate to AI systems – even if only on a pro rata basis – should be excreted.^[701] This is the only way to ensure that the right to be heard and the principles of fairness and integrity of arbitration proceedings are adequately considered.
229. In addition, every output from AI systems on which a decision is based should be checked by a human.^[702] This would allow the advantages of machine and human decision-making to be combined in individual cases.^[703] Naturally speaking, this does not rule out the possibility of using AI tools downstream, for example following a human search.^[704]
230. Moreover, it is particularly important to prematurely inform the (opposing) parties in a comprehensible and comprehensive manner about the use of AI tools, including its nature and scope. This should encompass the function as well as the intended type of use of the respective AI tool and concrete (significant) effects on the proceedings.^[705] Nothing else should apply to information about the complete prompt and the associated output of an AI processing.^[706] In particular, decision parameters and statistical bases should generally be disclosed, as well as the decision-relevant use of AI results outside a protocol that (may) actually influence an arbitrator’s understanding.
231. In accordance with the SVAMC guidelines, the use of ‘explainable AI’ functions is recommended.^[707] The aim is to make it comprehensible to human users how an AI system arrives at a certain result based on certain inputs. Such explanations are owed irrespective of any technical and cost limitations of explaining how complex AI systems work or of the general need for all participants in arbitration to exercise their own judgment independently and to be aware of potential bias that may be inherent in the outcome of AI systems.^[708] Accordingly, the due process principle requires arbitrators to independently and critically assess the reliability of AI information.^[709]
232. The guidelines also include obligations for the parties and their representatives: These include upstream due diligence obligations when using AI tools (eg, to check them in advance for application errors such as hallucinations)^[710] or the obligation to refrain from using (generative) AI if this is likely to jeopardize the integrity of the proceedings or the authenticity of evidence (eg, by means of deep fakes).^[711]
233. Finally, it must be ensured that the use of AI tools complies with the (legal) obligations to protect confidential information.^[712]
234. If AI is used to support arbitration, be it in researching and interpreting facts and the law or in applying the law to a concrete set of facts, these are to be classified as so-called high-risk AI systems within the meaning of Art 6(2), Annex III no 8 lit a) of the AI Act oft he European Union.^[713] This follows from the fact that decisions of arbitration tribunals are binding on the parties. Consequently, the providers and operators of such high-risk AI systems are subject to the special obligations of Art 9-17 AI Act.^[714]

4 Summary of the Main Results

4.1 Platform-Based Enforcement of Rights and Standards

4.1.1 Economization of Platform Procedures

235. Online platforms (provisionally) enforce state law as well as their own standards. By providing independent complaint procedures for arising disputes, usually in an easily accessible, cost- and time-saving manner, such disputes almost invariably no longer reach the state courts. This ‘absorption’ effect stems from the privatization of judicial, enforcement and regulatory tasks by private actors. What is more, the business model tailored by a platform to specific user (groups) is often continued in an ‘economization’ of internal procedural structures: Not infrequently, the result is a different treatment of (sometimes segregated) user groups: be it in terms of access to internal complaint procedures, be it with regard to the person or the technical (algorithmic) medium designated by a platform for making the decision. As a result, the principle of equal treatment of parties or participants with equivalent procedural roles, which is derived from national codes of (civil) procedure, is partly put into perspective. At the same time, the economization of platform proceedings is accompanied by adapted patterns of behavior on the part of rightsholders. In the context of copyright law, such adaptation processes are expressed, for example, in an increasing monetization of protected content, thereby replacing reactive law enforcement usually aimed at deleting infringing content. In this respect, the (procedural) role of online platforms is advancing from ‘mere’ mediators of sanctions to mediators of economic revenue growth.

4.1.2 Merging of State and Private Enforcement Interests

236. Another line of development emerges in parallel: the merging of state and private enforcement interests. Which proves to be problematic whenever (and due to) online platforms not (clearly) separating the enforcement of state-granted rights on the one hand and their own private standards on the other. By resorting to platforms in order to ostensibly enforce the standards of those private actors without disclosing and justifying the actual public enforcement interests, state authorities run the risk of circumventing the rule of law. In addition, private platform standards are often interpreted in line with a specific pre-understanding or regulatory calculation of a state. As a consequence, the guiding function and recognizability of state law will be weakened. Compared to this, internal complaint procedures offer little effective protection, especially as online platforms do not (yet) provide any specific procedures or procedural rules against forms of state involvement in a platform-based enforcement of rights and standards. Last but not least, by instrumentalizing large international online platforms, government agencies hope to enforce national law de facto with transnational effect, as illustrated by the global application of the fair use principle inherent in US copyright law.

4.1.3 Potential Effects of the Use of Artificial Intelligence

237. The use of AI supports and reinforces such effects. In platform-based dispute resolution, as in arbitration or conciliation, it harbors the risk of opaque decision bases (the so-called black box problem). Thus, AI regularly prevents individual steps in the decision-making process from being reconstructed. It is the independence of the output from the input, often in conjunction with feedback loops, that favors the phenomenon of algorithmic discrimination. Furthermore, AI is technically incapable of showing empathy – the cornerstone of interpersonal communication – nor can it make cognitive or (spontaneous) judgment-based or value-related decisions. The same applies to the (lacking) ability to grasp highly context-sensitive content in a legally correct manner. In this light, it is hardly possible to justify (enforcement) decisions at all. In addition, AI’s strict reference to the past makes it impossible to develop the law in individual cases. All of these potential consequences considerably affect the effective legal defense of both platform users and rightsholders. At the same time, rapid technical progress offers the possibility of ever greater flexibility and individualization of platform procedures.

4.1.4 Proceduralization of Private Rights and Standards Enforcement

238. To such dangers the (increased) proceduralization of private rights and standards enforcement offers a solution. This is because procedural structures fulfil an independent and necessary function in order to ensure effective legal protection in the context of private, platform-based rights enforcement: Only then do they enable the addressee of enforcement, but also affected rightsholders, especially in particularly dangerous situations, to effectively realize their rights and thus to provisionally balance them in the context of uncertain legal and factual situations in a less faulty manner. The inclusion of platforms under liability law alone does not do justice to this regulatory objective: it is neither geared towards such a balancing of interests nor does it ask about the procedural conditions of possibility and the implementation of duties of disclosure.
239. Recently, state regulatory acts have taken a first, corresponding approach. The European Union’s Digital Services Act serves as the inspiration. Hereby procedural (minimum) standards, essentially borrowed from the guarantees of state (civil) procedure, are being put in place. These include, inter alia, the right to equal access to proceedings, the obligation of a platform operator to justify (enforcement) decisions as well as the right to an effective complaint. In the interests of platform users, these rights particularly aim to guarantee the right to be heard and an effective legal defense and, therefore, to balance structural power and information imbalances within platform structures. Above this, a multipolar, more differentiated conflict resolution system is emerging that does not only include online platforms, rightsholders and users: In order to promote effective law enforcement, platform procedures also involve third parties such as (trusted) whistleblowers.
240. Nevertheless, there is still a need for further regulation: For instance, the objective of more closely dovetailing a platform’s internal complaint procedures with procedures for independent out-of-court dispute resolution has not yet been sufficiently implemented. This applies all the more in light of a structural bias of platforms, especially when enforcing their own standards. In order to compensate for the limited redeemability of substantively correct private legal enforcement, internal complaint procedures – all the more in view of (automated) enforcement of a platform’s own standards – should be supplemented by an out-of-court dispute settlement subject to a subsequent state court decision. The procedure should be freely accessible to affected users, rightsholders and whistleblowers and binding for platforms.
241. The model of provisional legal protection as provided for by the national codes of civil procedure is a suitable starting point for proceduralizing platform-based legal enforcement: In particular, this applies with regard to a platform’s reduced duties to clarify the facts and assumptions of probability underlying a platform’s decision, but also to a hearing of addressees downstream of urgent enforcement measures. In order to assess a platform’s liability risks more reliably, clearly defined procedural (reaction) obligations should take the place of case-by-case duties of disclosure; the substantive standard of review should be simplified and, for example, substantiated by legal presumptions or a typification of (un)permitted user behavior.
242. Equally a stronger proceduralization of platform-based decision-making has only been implemented to a limited extent through government regulatory acts. In addition to requirements for (better) qualification of platform decision-makers and fundamental decision-making standards under the rule of law for platform-based legal enforcement (such as the obligation to act carefully, objectively, without arbitrariness and proportionately), this also concerns issues of consistency and comparability of platform decisions, which are often made more difficult, if not prevented, by non-transparent decision-making standards. Rules on the scaling of decisions, ie, when platforms transfer specific interpretation standards derived from ‘model cases’ and deemed to be of ‘high quality’ to comparable situations are also required. Better protection against distortion effects, which are particularly associated with context-sensitive disputes and often have a transnational reach, is promised by, among other things, increased transparency obligations, further a stronger link between scaling decisions and external objective review standards and independent review bodies, as well as effective downstream legal protection.
243. Transparency obligations of platforms are a key component of proceduralization. In this context, they fulfill both an auxiliary procedural function and a function of institutional control. By the means of periodic transparency reports or a right of access for research purposes, they are intended to ensure preventive public control of platforms – whether by authorities (eg, to protect competition and consumer interests) or by academia, the media and non-governmental organizations. Then again, clear provisions in the general terms and conditions of online platforms regarding the intended forms of sanctions should make these predictable for individual platform users, especially when using AI: Corresponding information obligations should cover, among other things, the course of platform procedures, whether automated systems are used or whether and on what basis platforms have taken measures at the instigation of a state authority. Last but not least, specific decisions made by online platforms, for instance, on the removal and blocking of access or the downgrading of content, should be stored in a publicly accessible database. Ensuring procedural transparency is therefore a central procedural principle of platform-based law enforcement.

4.1.5 The Regulatory Dimension: Interaction Between (Individual) Rule Enforcement and (Collective) Setting of Rules

244. Standardization of platform-based rule enforcement can also be observed. This is the result of feedback from individual platform decisions to a platform’s AI systems. At the same time, the ongoing implementation of platform-specific standards influences their dynamic, regulatory development.
245. This ‘collectivizing’ effect is reinforced by so-called platform courts, ie, non-governmental, court-like dispute resolution bodies such as the Facebook Oversight Board. Such courts will not only have the task of settling individual disputes (usually of particular relevance) and continuously harmonizing decision-making and interpretation of a platform’s community standards. Platform courts at once shape a private process of rule-setting, essentially following a ‘procedural concept of (self-)regulation’. This process of norm-building is (partly) beyond state control of the framework conditions, while at the same time being communicatively tied back to the public. The aim is twofold: on the one hand, to ensure more effective protection of individual rights in a multitude of (parallel) cases and, on the other, to enable a continuously adapted, by its nature mostly rudimentary-experimental rule building through and within the framework of procedural structures. In this way, interpretation (contentious review), setting and enforcement of private standards ultimately amalgamate into ‘one power’.
246. Such a concentration of power of online platforms can be countered from the perspective of various areas of law. As far as procedural solutions are concerned, out-of-court dispute settlement (cf Art 21 of the EU DSA Regulation) opens up an incidence check of a platform’s T&Cs and its moderation guidelines concretizing them. Reference point of control is their conformity with the procedural and contractual requirements of relevant EU legal acts (cf Digital Services Act; P2B Regulation). From the perspective of such legislative overlay, platform T&Cs thus (de facto) assume a regulatory function. This indirect control of a platform’s own standards is flanked by opportunities for direct state influence, primarily in the form of regulating internal platform procedures. This approach finds paradigmatic expression in the Digital Services Act in the binding specification of procedural (minimum) requirements.

4.2 Use of Artificial Intelligence in Mediation, Dispute Resolution and Arbitration Proceedings

247. In contrast to a platform-based enforcement of rights and standards, the use of AI in mediation, dispute resolution and arbitration proceedings has so far only fulfilled an upstream auxiliary function. It is primarily aimed at increasing procedural efficiency, ie, essentially at saving time and costs in the interests of the parties to the dispute and the independent arbitrator, conciliator, or mediator. The scope of application of AI ranges from providing procedural information, including simple legal information, automated document analyses and other forms of case management to the preparation of settlement and decision proposals (predictive tools). A ‘supra-individual’ standardization of dispute resolution is beginning to emerge, insofar as companies geared towards dispute resolution store the data of all disputes, which are then available to them as data sets for each new legal dispute pending before such companies.
248. The risks associated with these areas of AI application are similar to those of platform-based law and standards enforcement, for example in the form of the black box problem or algorithmic discrimination. However, these risks have a less serious impact on arbitration proceedings as well as in mediation and dispute resolution. This is because fully automated or AI-based proceedings without a human decision-maker are currently neither technically possible nor desirable in terms of legal policy. Only forms of ‘hybrid decision-making’ are already in use. The de facto settlement of disputes, which is characteristic of a platform-based enforcement of rights and standards, does not occur in mediation, dispute resolution or arbitration proceedings either.

---

Abbreviations and Acronyms

ADR	Alternative Dispute Resolution
Art	Article/Articles
BGB	Bürgerliches Gesetzbuch (Civil Code) [Germany]
BGH	Bundesgerichtshof (Federal Court of Justice) [Germany]
BVerfG	Bundesverfassungsgericht (Federal Constitutional Court) [Germany]
cf	confer (compare)
ch	chapter
CJEU	Court of Justice of the European Union
ECLI	European Case Law Identifier
ECtHR	European Court of Human Rights
DMCA	Digital Millennium Copyright Act [USA]
DSA or DSA Regulation	Digital Services Act [European Union]
ECG	E-Commerce-Gesetz (E-Commerce Act) [Austria]
ed	editor/editors
edn	edition/editions
eg	exempli gratia (for example)
etc	et cetera
EU	European Union
EUR	Euro
f/ff	following
fn	footnote (external, ie, in other chapters or in citations)
GSR	General Secondary Response
GG	Grundgesetz (Fundamental Law) [Germany]
GDPR	General Data Protection Regulation [European Union]
GTC	General terms and conditions
ibid	ibidem (in the same place)
ie	id est (that is)
KoPlG	Kommunikationsplattformengesetz (Communications Platforms Act) [Austria]
n	footnote (internal, ie, within the same chapter)
NetzDG	Netzwerkdurchsetzungsgesetz (Network Enforcement Act) [Germany]
no	number/numbers
öABGB	Allgemeines Bürgerliches Gesetzbuch (General Civil Code) [Austria]
öUrhG	Urheberrechtsgesetz (Copyright Act) [Austria]
P2B Regulation	Regulation (EU) 2019/1150 of 20 June 2019 on promoting fairness and transparency for business users of online intermediary services
PACT or PACT Act	Platform Accountability and Consumer Transparency Act [USA]
para	paragraph/paragraphs
pt	part
RDG	Rechtsdienstleistungsgesetz (Legal Services Act) [Germany]
S.D. Cal.	District Court for the Southern District of California [USA]
sec	Section/Sections
supp	supplement/supplements
TMG	Telemediengesetz (Telemedia Act) [Germany]
trans/tr	translated, translation/translator
UGC	User-generated content
UK	United Kingdom
UNIDROIT	Institut international pour l'unification du droit privé (International Institute for the Unification of Private Law)
UrhDaG	Urheberrechts-Diensteanbieter-Gesetz (Act on Copyright Content Sharing Service Providers) [Germany]
UrhG	Urheberrechtsgesetz (Copyright Act) [Germany]
US/USA	United States of America
U.S.C.	United States Code
USD	United States Dollar
v	versus
vol	volume/volumes
VSBG	Verbraucherstreitbeilegungsgesetz (Consumer Dispute Resolution Act) [Germany]
ZPO	Zivilprozessordnung (Code of Civil Procedure) [Germany]
***	***

---

Legislation

International/Supranational

Charter of Fundamental Rights of the European Union, OJ C 364, 18 December 2000, 1.

Comission and Parliament (EU), Synopsis AI Act, P9_TA(2023)0236, https://www.europarl.europa.eu/doceo/document/TA-9-2023-0236_EN.pdf accessed 31 December 2023.

Commission Recommendation (EU) 2018/334 of 1 March 2018 on measures to effectively tackle illegal content online, OJ L 62, 6 March 2018, 50.

Directive 2000/31/EC of 8 June 2000 on certain legal aspects of information society services, in particular electronic commerce, in the Internal Market (‘Directive on electronic commerce’), OJ L 178, 17 July 2000, 1.

Directive 2001/29/EC of 22 May 2001 on the harmonization of certain aspects of copyright and related rights in the information society, OJ L 167, 22 June 2001, 10.

Directive 2004/48/EC of 29 April 2004 on the enforcement of intellectual property rights, OJ L 157, 30 April 2004, 45.

Directive 2008/52/EC of the European Parliament and of the Council of 21 May 2008 on certain aspects of mediation in civil and commercial matters, OJ L 136, 24 May 2008, 3.

Directive 2013/11/EU of 21 May 2013 on alternative dispute resolution for consumer disputes and amending Regulation (EC) No 2006/2004 and Directive 2009/22/EC (Directive on consumer ADR), OJ L 165, 18 June 2013, 63.

Directive 2019/790/EU of 17 April 2019 on copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC, OJ L 130, 17 May 2019, 92.

ICC Arbitration Rules 2021.

Regulation (EU) 2019/1150 of 20 June 2019 on promoting fairness and transparency for business users of online intermediation services, OJ L 186, 11 July 2019, 57.

Regulation (EU) 2021/784 of 29 April 2021 on addressing the dissemination of terrorist content online, OJ L 172, 17 May 2021, 79.

Regulation (EU) 2022/2065 of 19 October 2022 on a Single Market For Digital Services and amending Directive 2000/31/EC (Digital Services Act; hereinafter: DSA Regulation), OJ L 277, 27 October 2022, 1.

Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act), OJ L, 12 July 2024.

UNCITRAL Model Law on International Commercial Arbitration (1985).

***

National

Arbitration Rules of the Netherlands Arbitration Institute (Schiedsordnung Nederlands Arbitrage Instituut).

Austrian Communications Platforms Act (Kommunikationsplattformengesetz: KoPlG).

Austrian Copyright Act (Urheberrechtsgesetz: öUrhG).

Austrian General Civil Code (Allgemeines Bürgerliches Gesetzbuch: öABGB).

Austrian E-Commerce Act (E-Commerce-Gesetz: ECG).

English Arbitration Act 1996.

French Code of Civil Procedure (Code de procédure civile).

French Law of 23 March 2019 (Loi de programmation 2018-2022 et réforme pour la justice) No 2019/222.

German Act on Copyright Content Sharing Service Providers (Urheberrechts-Diensteanbieter-Gesetz: UrhDaG).

German Fundamental Law (Grundgesetz: GG).

German Civil Code (Bürgerliches Gesetzbuch: BGB).

German Code of Civil Procedure (Zivilprozessordnung: ZPO).

German Consumer Dispute Resolution Act (Verbraucherstreitbeilegungsgesetz: VSBG).

German Copyright Act (Urheberrechtsgesetz: UrhG).

German Legal Services Act (Rechtsdienstleistungsgesetz: RDG).

German Network Enforcement Act (Netzwerkdurchsetzungsgesetz: NetzDG).

German Telemedia Act (Telemediengesetz: TMG).

Italian Code of Civil Procedure (Codice di procedura civile: CPC).

Spanish Arbitration Act (Lex 60/2003 de 23 de diciembre, de Arbitraje).

Title 17 of the United States Code – Copyright Law (17 U.S.C.).

Title 47 of the United States Code – Telecommunications (47 U.S.C.).

Turkish International Arbitration Law (Uluslararası Tahkim Kanunu).

United States of America Platform Accountability and Consumer Transparency Act, 116th Cong. § 5(2) (2020) (PACT Act).

United States of America Digital Millennium Copyright Act (DMCA).

***

---

Cases

International/Supranational

CJEU, 23 March 2010, C-236/08 and C-238/08 – Google France, ECLI:EU:C:2010:159.

CJEU, 12 July 2011, C‑324/09 – L’Oréal v eBay, ECLI:EU:C:2011:474.

CJEU (Grand Chamber), 24 September 2019, C-507/17 – Google (Spatial scope of delisting), ECLI:EU:C:2019:772.

CJEU, 3 October 2019, C-18/18 – Glawischnig-Piesczek, ECLI:EU:C:2019:821.

CJEU, 26 April 2022, C‑401/19 – Poland v Parliament and Council, ECLI:EU:C:2022:297.

European General Court, 27 September 2023, T-367/23 – Amazon Services Europe v Commission, ECLI:EU:T:2023:589.

Oversight Board, 28 January 2021, 2020-005-FB-UA – Nazi Quote.

Oversight Board, 5 May 2021, 2021-001-FB-FBR – Former President Trump's suspension.

Oversight Board, 8 July 2021, 2021-006-IG-UA – Ocalan's Isolation.

Oversight Board, 14 September 2021, 2021-009-FB-UA – Shared Al Jazeera Post.

Oversight Board, 27 September 2021, 2021-010-FB-UA – Colombia Protests.

Oversight Board, 17 June 2022, 2022-001-FB-UA – Knin Cartoon.

Oversight Board, 15 September 2022, 2022-005-FB-UA – Mention of the Taliban in News Reporting.

Oversight Board, 22 November 2022, 2022-007-IG-MR – UK Drill Music.

Oversight Board, 6 December 2022, 2021-002-FB-PAO – Meta's Cross-Check Program.

Oversight Board, 14 December 2022, 2022-012-IG-MR – India Sexual Harassment Video.

Oversight Board, 14 December 2022, 2022-011-IG-UA – Video after Nigeria Church Attack.

Oversight Board, 9 January 2023, 2022-013-FB-UA – Iran Protest Slogan.

Oversight Board, 9 March 2023, 2022-014-FB-MR – Sri Lanka Pharmaceuticals.

Oversight Board, 18 December 2023, 2023-054-FB-UA, 2023-055-FB-UA, 2023-056-FB-UA, 2023-057-FB-UA – Goebbels Quote.

***

National

BGH (Germany), 17 August 2011, I ZR 57/09 – Stiftparfüm, BGHZ 191, 19 = (2011) 113(11) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 1038.

BGH (Germany), 25 October 2011, VI ZR 93/10 – Blog Eintrag, BGHZ 191, 219 = (2012) 65(3) NJW (Neue Juristische Wochenschrift) 148 = (2012) 114(3) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) GRUR 2012, 311.

BGH (Germany), 27 October 2011, I ZR 131/10 – regierung-oberfranken.de, (2012) 65(31) NJW (Neue Juristische Wochenschrift) 2279.

BGH (Germany), 12 July 12, 2012, I ZR 18/11 – Alone in the dark, BGHZ 194, 339 = (2013) 66(11) NJW (Neue Juristische Wochenschrift) 784.

BGH (Germany), 18 June 2015, I ZR 74/1 – Liability for Hyperlink, BGHZ 206, 103 = (2016) 69(11) NJW (Neue Juristische Wochenschrift) 804.

BGH (Germany), 1 March, 2016, VI ZR 34/15 – Ärztebewertungsportal III (jameda.de), BGHZ 209, 139 = (2016) 65(3) NJW (Neue Juristische Wochenschrift) 2106.

BGH (Germany), 22 November 2017, VIII ZR 83/16, (2018) 71(8) NJW (Neue Juristische Wochenschrift) 537.

BGH (Germany), 22 November 2017, VIII ZR 213/16, (2018) 21(3) MMR (Multimedia und Recht) 156.

BGH (Germany), 29 July 2021, III ZR 179/20, (2021) 74(43) NJW (Neue Juristische Wochenschrift) 3179.

BGH (Germany), 29 July 2021, III ZR 192/20, (2021) 25(11) ZUM-RD (Zeitschrift für Urheber- und Medienrecht – Rechtsprechungsdienst) 612.

BVerfG (Germany), 3 February 1959, 2 BvL 10/56, BVerfGE 9, 137 = (1959) 12(21) NJW (Neue Juristische Wochenschrift) 931.

BVerfG (Germany), 8 August 1978, 2 BvL 8/77, BVerfGE 49, 89 = (1979) 32(8) NJW (Neue Juristische Wochenschrift) 359.

BVerfG (Germany), 20 April 1982, 2 BvL 26/81, BVerfGE 60, 253 = (1982) 35(43) NJW (Neue Juristische Wochenschrift) 1982, 2425.

District Court for the Southern District of California (USA), United States v Green, 857 F. Supp. 2d 1015, 1018 (S.D. Cal. 2012).

***

---

Bibliography

Achleitner R, ‘The Fight against Geo-Blocking – A Never Ending Story? Policy Paper on Geo-Blocking’ https://ssrn.com/abstract=4246896 or http://dx.doi.org/10.2139/ssrn.4246896 accessed 31 December 2023.

Adolphsen J, ‘Der Zivilprozess im Wettbewerb der Methoden’ (2017) 48(4) BRAK Mitteilungen 147.

Althammer C, ‘Alternative Streitbeilegung im Internet’ in F Faust and H-B Schäfer (ed), Zivilrechtliche und rechtsökonomische Probleme des Internet und der künstlichen Intelligenz (Mohr Siebeck 2019) 249.

Ameln F von, ‘Führen und Entscheiden unter Unsicherheit‘ (2021) 52(4) GIO (Gruppe. Interaktion. Organisation. Zeitschrift für Angewandte Organisationspsychologie) 567.

Andrea R, ‘No Safe Harbor: YouTube’s Content ID and Fair Use’ (2020) Boston College Intellectual Property & Technology Forum 1.

Anzinger H M, ‘10 Jahre Modria – KMS und Online-Mediation auf dem Weg zur Digitalisierung der Justiz – Teil 1‘ (2021) 24(2) ZKM (Zeitschrift für Konfliktmanagement) 53.

Anzinger H M, ‘10 Jahre Modria – KMS und Online-Mediation auf dem Weg zur Digitalisierung der Justiz – Teil 2’ (2021) 24(3) ZKM (Zeitschrift für Konfliktmanagement) 84.

AIRBNB, ‘Scoring the user to prevent “suspicious” activity before it occurs: What Does It Mean When Someone’s ID Has Been Checked?’, https://www.airbnb.com/help/article/2356/what-does-it-mean-when-someones-id-has-been-checked accessed 31 December 2023.

Appelman N and Leerssen P, ‘On “Trusted” Flaggers’ (2022) 24 Yale Journal of Law & Technology 452.

Arbitrator Intelligence, ‘Arbitrator Intelligence Database’, https://arbitratorintelligence.vercel.app/ accessed 31 December 2023.

Aschauer C, ‘Automated Decision-Making and Artificial Intelligence (AI) in Arbitration’ in C Leyens, I Eisenberger and R Niemann (ed), Smart Regulation (Mohr Siebeck 2021) 130.

Askani A, Private Rechtsdurchsetzung bei Urheberrechtsverletzungen im Internet (Nomos 2021).

Bambauer D E, ‘Against Jawboning’ (2015) 100(1) MINN. L. REV. 51.

Barocas S and Self A D, ‘Big Data’s Disparate Impact’ (2016) 104(3) California Law Review 671 (2016).

Barona Vilar S, ‘Effizienszsteigerung und Suche nach Beschleunigung von Schiedsverfahren im Spannungsfeld von Mythos. Sublimierung und Vierter Industrieller Revolution (4.0)’ (2018) 23 ZZPInt (Zeitschrift für Zivilprozess International) 295.

Barton H, ‘Rebooting Justice’ (2018) 44(4) Law Practice 32.

Bar-Ziv S and Elkin-Koren N, ‘Behind the Scenes of Online Copyright Enforcement: Empirical Evidence on Notice & Takedown’ (2018) 50(2) Connecticut Law Review 339.

Berberich M and Conrad A, ‘§ 30 Plattformen und KI’ in M Ebers, CA Heinze and B Steinrötter (ed), Künstliche Intelligenz und Robotik (Beck 2020) 930.

Berberich M, ‘§ 5 Sorgfaltspflichten, Moderationsverfahren und prozedurale Fairness’ in B Steinrötter (ed), Europäische Plattformreguliereng (Nomos 2023) 126.

Bizikova L, Hancock P, Jewell D and Sherr I, ‘IA Meets AI – Rise of the Machines’, <https://dailyjus.com/legal-tech/2023/10/ia-meets-ai-rise-of-the-machines> accessed 31 December 2023.

Bloch-Wehba H, ‘Automation in Moderation’ (2020) 53(1) Cornell International Law Journal 41.

Bloch-Wehba H, ‘Global Platform Governance: Private Power in the Shadow of the State’ (2019) 72(1) SMU Law Review 27.

Boll-Kempelmann C, ‘Datenschutz und das Beweisverfahren im Schiedsverfahren’ (2022) 20(5) SchiedsVZ (Zeitschrift für Schiedsverfahren) 241.

Bomhard D and Merkle M, ‘Regulation of Artificial Intelligence’ 2021 (6) EuCML (Journal of European Consumer and Market Law) 257.

Brazil W, ‘Informalism and Formalism in the History of ADR in the United States’ in J Zekoll, M Bälz and I Amelung (ed), Formalisation and Flexibilisation in Dispute Resolution (Brill 2014) 250.

Brechmann B, Legal Tech und das Anwaltsmonopol (Mohr Siebeck 2021).

Breiman L, ‘Random Forests’ (2001) 45(1) Machine Learning 5.

Bridy A, ‘Intellectual Property’ in Keller D (ed), Law, Borders, and Speech: Proceedings and Materials (The Center for Internet and Society 2017) 9.

Bünnau P von, ‘Künstliche Intelligenz im Recht’ in S Breidenbach and F Glatz (ed), Rechtshandbuch Legal Tech (2nd edn, Beck/Manz 2021) 71.

Bull L and Steffek F, ‘The Decoding of Legal Conflicts’ (2018) 21(5) ZKM (Zeitschrift für Konfliktmanagement) 165.

Burk D L, ‘Algorithmic Fair Use’ (2019) 86(2) University of Chicago Law Review 283.

Burrell J, ‘How the machine “thinks”: Understanding opacity in machine learning algorithms’ (2016) (1) Big Data & Society 3, https://doi.org/10.1177/2053951715622512 accessed 31 December 2023.

Busch C, ‘Mehr Fairness und Transparenz in der Plattformökonomie?’ (2019) 121(8) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 788.

Calliess C and Ruffert M (ed), EUV/AEUV (6th edn, Beck 2022).

Casey B et al, ‘Rethinking Explainable Machines: The GDPR's “Right to Explanation” Debate and the Rise of Algorithmic Audits in Enterprise’ (2019) 34(1) BERKELEY TECH. L.J. 145.

Castets-Renard C, ‘Algorithmic Content Moderation on Social Media in EU Law: Illusion of Perfect Enforcement’ (2020) (2) University of Illinois Journal of Law, Technology & Policy 283.

Cellan-Jones R, ‘The robot lawyers are here’, <https://www.bbc.com/news/technology-41829534> accessed 31 December 2023.

Cervenka A and Schwarz P, ‘Data Protection in Arbitration Proceedings’ (2020) 18(2) SchiedsVZ (Zeitschrift für Schiedsverfahren) 78.

Chang B, ‘From Internet Referral Units to International Agreements: Censorship of the Internet by the UK and EU’ (2018) 49(2) Columbia Human Rights Law Review 114.

Chander A, ‘The Racist Algorithm?’ (2017) 115(6) Michigan Law Review 1023.

Chargeflow, ‘Paypal Dispute automation’, https://www.chargeflow.io/paypal-dispute-automation accessed 31 December 2023.

Citron D K, ‘Technological due process’ (2008) 85(6) Washington University Law Review 1249.

Citron D K and Pasquale F, ‘The Scored Society: due process for automated predictions’ (2014) 89(1) WASH. L. REV. 1.

Civil resolutional tribunal, ‘Civil resolutional tribunal’, https://civilresolutionbc.ca/solution-explorer/ accessed 31 December 2023.

Cohen J E, Between Truth and Power: The Legal Constructions of Informational Capitalism (Oxford University Press 2019).

Cohen P, ‘Bytes and Prejudice’ (2015) 1(1) Journal of Technology in International Arbitration 57.

Conrad A and Nolte G, ‘Schrankenbestimmungen im Anwendungsbereich des UrhDaG’ (2021) 65 ZUM (Zeitschrift für Urheber- und Medienrecht) 111.

Cooper S, Rule C and Del Duca L, ‘From Lex Mercatoria to Online Dispute Resolution’ (2011) 43 Uniform Commercial Code Law Journal, Penn State Legal Studies Research Paper No 09/2011.

Deckenbrock C and Henssler M, Rechtsdienstleistungsgesetz (5th edn, Beck 2021).

Deichsel T, Digitalisierung der Streitbeilegung (Nomos 2022).

Deichsel T, ‘Verbraucherschlichtungsstellen – Ein Anwendungsfeld für Legal Tech?‘ (2020) 35(8) VuR (Verbraucher und Recht) 283.

Denga M, ‘Platform Regulation by European Values: On the Binding of Opinion Platforms to EU Fundamental Rights’ (2021) 56(5) EuR (Europarecht) 569.

Deusch F and Eggendorfer T, ‘IT-Sicherheit’ in J Taeger and J Pohle (ed), Computerrechts-Handbuch (38 th edn, Beck 2023) ch 50.1.

Dewey C, ‘98 personal data points that Facebook uses to target ads to you’, The Washington Post, 19 August 2016, https://www.washingtonpost.com/news/the-intersect/wp/2016/08/19/98-personal-data-points-that-facebook-uses-to-target-ads-to-you/ accessed 31 December 2023.

Dispute Resolution Data, ‘Dispute Resolution Database’, https://www.disputeresolutiondata.com/ accessed 31 December 2023.

Disco, ‘Disco Powerful AI and Analytics’, https://www.csdisco.com/offerings/ediscovery/features-ai accessed 31 December 2023.

Douek E, ‘Content Moderation as Systems Thinking’ (2022) 136(2) Harvard Law Review 528.

Douek E, ‘How Much Power Did Facebook Give Its Oversight Board?’, <https://www.lawfaremedia.org/article/how-much-power-did-facebook-give-its-oversight-board> accessed 31 December 2023.

Douek E, ‘The Oversight Board Moment You Should've Been Waiting For, Lawfare’, <https://www.lawfaremedia.org/article/oversight-board-moment-you-shouldve-been-waiting-facebook-responds-first-set-decisions> accessed 31 December 2023.

Drexl J, ‘Bedrohung der Meinungsvielfalt durch Algorithmen’ (2017) 61(7) ZUM (Zeitschrift für Urheber- und Medienrecht) 529.

Drupal, ‘a2J Author’, https://www.a2jauthor.org/ accessed 31 December 2023.

Dubois E and Blank G, ‘The echo chamber is overstated: the moderating effect of political interest and diverse media’ (2018) 21(5) Information, Communication & Society 729.

Ebrevia, ‘DraftPro’, https://www.dfinsolutions.com/products/ebrevia accessed 31 December 2023.

Eghbariah R and Metwally A, ‘Informal Governance: Internet Referral Units and the Rise of State Inter-pretation of Terms of Service’ (2021) 23 Yale J.L. & Tech. 545.

Eidenmüller H and Wagner G, Law by Algorithm (Mohr Siebeck 2021).

Eifert M et al, Netzwerkdurchsetzungsgesetz in der Bewährung (Nomos 2020).

Elkin-Koren N, ‘After twenty years: revisiting copyright liability of online intermediaries’ in S Frankel and D Gervais (ed), The Evolution and Equilibrium of Copyright in the Digital Age (Cambridge University Press 2014) 29.

Elkin-Koren N, ‘Contesting algorithms: Restoring the public interest in content filtering by artificial intelligence’ (2020) 7(2) Big Data & Society 1, https://journals.sagepub.com/doi/epub/10.1177/2053951720932296 accessed 31 December 2023.

Elkin-Koren N and Perel M, ‘Separation of Functions for AI: Restraining Speech Regulation by Online Platforms’ (2020) 24(3) Lewis & CLARK L. REV. 857.

Engert A, ‘Digitale Plattformen’ (2018) 218(2-4) AcP (Archiv für die civilistische Praxis) 304.

Engstrom E and Feamster N, The Limits of Filtering: A Look at the Functionality & Shortcomings of Content Detection Tools (Engine 2017), https://static1.squarespace.com/static/571681753c44d835a440c8b5/t/58d058712994ca536bbfa47a/1490049138881/FilteringPaperWebsite.pdf accessed 31 December 2023.

European Commission, ‘Staff Working Document Impact Assessment, Accompanying the document Pro-posal for a Regulation of the European Parliament and of the Council on preventing the dissemina-tion of terrorist content online’ (12 September 2018) SWD (2018), 408 final.

European Commission, ‘Tackling online disinformation: Commission proposes an EU-wide Code of Practice’ (26 April 2018), https://ec.europa.eu/commission/presscorner/detail/en/IP_18_3370 accessed 31 December 2023.

Europol, ‘2018 Consolidated Annual Activity 44’ (23 May 2019), https://www.europol.europa.eu/cms/sites/default/files/documents/consolidated_annual_activity_report_2018.pdf accessed 23 December 2023.

Europol, ‘EU Internet Referral Unit – EU IRU’, https://www.europol.europa.eu/about-europol/european-counter-terrorism-centre-ectc/eu-internet-referal-unit-eu-iru accessed 31 December 2023.

Facebook Oversight Board, ‘Oversight Board Bylaws’, https://www.oversightboard.com/wp-content/uploads/2024/03/Oversight-Board-Bylaws.pdf accessed 31 December 2023.

Facebook Oversight Board, ‘Oversight Board Charter’, https://about.fb.com/wp-content/uploads/2019/09/oversight_board_charter.pdf accessed 31 December 2023.

Fiala F and Husovec M, ‘Using Experimental Evidence to Improve Delegated Enforcement’ (2022) 71 International Review of Law and Economics 106079 and (2018) 28 Forthcoming TILEC Discussion Paper no 2018-028.

Fireflies, ‘Fireflies AI’, https://firefliesai/ accessed 31 December 2023.

Fries M, ‘Erfüllung von Geldschulden über eigenwillige Zahlungsdienstleister’ (2018) 33(4) VuR (Verbraucher und Recht) 123.

Fries M, ‘Legal Tech im Schiedsverfahren’ in R Wilhelmi and M Stürner (ed), Mehrparteienschiedsverfahren. Unter besonderer Berücksichtigung gesellschaftsrechtlicher Streitigkeiten (Springer 2021) 85.

Fries M, ‘PayPal Law und Legal Tech – Was macht die Digitalisierung mit dem Privatrecht?’ (2016) 69(39) NJW (Neue Juristische Wochenschrift) 2860.

Fries M, Verbraucherrechtsdurchsetzung (Mohr Siebeck 2016).

Fritz G, Prantl D, Leinwather N and Hofer M, ‘Data Protection in International Arbitration Proceedings’ (2019) 17(6) SchiedsVZ (Zeitschrift für Schiedsverfahren) 301.

Frosio G, ‘Algorithmic Enforcement Online’ in P Torremans (ed), Intellectual Property and Human Rights (4th edn, Kluwer Law International 2020) 24.

Gerdemann S and Spindler G, ‘Das Gesetz über digitale Dienste (Digital Services Act) (Part 2) – Die Regelungen für Online-Plattformen sowie sehr große Online-Plattformen und Suchmaschinen’ (2023) 125(3) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 115.

Gielen N and Uphues S, ‘Digital Markets Act und Digital Services Act’ (2021) 32(14) EuZW (Europäische Zeitschrift für Wirtschaftsrecht) 627.

Gillespie T, ‘The Relevance of Algorithms’ in T Gillespie, P J Boczkowski and K A Foot (ed), Media Technologies (The MIT Press 2014) 188.

Gläßer U, ‘Mediation und Digitalisierung’ in T Riehm and S Dörr (ed), Digitalisierung und Zivilverfahren (De Gruyter 2023) 529.

Global Arbitration Review (GAR), ‘Database’, https://globalarbitrationreview.com/tools/arbitrator-research-tool accessed 31 December 2023.

Glogowski M, Plattformbedingungen (Mohr Siebeck 2022).

Google, ‘Asset erstellen’, https://support.google.com/youtube/answer/3011552?hl=de&ref_topic=3011550 accessed 31 December 2023.

Google, ‘Transparency Reports’, <https://transparencyreport.google.com/> accessed 31 December 2023.

Google, ‘Transparency Report for YouTube Platform for January to June 2023’ <https://transparencyreport.google.com/netzdg/youtube?hl=de> accessed 31 December 2023.

Google, ‘How Google Fights Piracy’, https://www.youtube.com/watch?v=QuOn93KPIr4 accessed 31 December 2023.

Google, ‘Counter notification’ https://support.google.com/youtube/answer/6005919?hl=de&ref_topic=9282678 accessed 31 December 2023.

Gorwa R, Binns R and Katzenbach C, ‘Algorithmic content moderation’ (2020) 7(1) Big Data & Society 1, https://journals.sagepub.com/doi/epub/10.1177/2053951719897945 accessed 31 December 2023.

Gray J E and Suzor N P, ‘Playing with machines: using machine learning to understand automated copyright enforcement at scale’ (2020) 7(1) Big Data & Society 1, https://journals.sagepub.com/doi/epub/10.1177/2053951720919963 accessed 31 December 2023.

Greger R and Stubbe C, Schiedsgutachten (Beck 2007).

Greger R, ‘Recht der alternativen Konfliktlösung’ in R Greger, H Unberath and F Steffek (ed), Recht der alternativen Konfliktlösung (2nd edn, Beck 2016) 270.

Grimmelmann J, ‘The Virtues of Moderation’ (2015) 17 YALE J.L. & TECH. 42.

Grosse Ruse-Kahn H, ‘Automated Copyright Enforcement Online: From Blocking to Monetization of User-Generated Content’ PIJIP Research Paper Series 51, https://digitalcommons.wcl.american.edu/research/51 accessed 31 December 2023.

Grosse Ruse-Kahn H, ‘Global Content Protection through Automation’ (2018) 49(9) IIC (International Review of Intellectual Property and Competition Law) 1017.

Gsell B, ‘Die Umsetzung der Richtlinie über alternative Streitbeilegung Juristisches Fachwissen der streitbeilegenden Personen und Rechtstreue des Verfahrensergebnisses’ (2015) 128(2) ZZP (Zeitschrift für Zivilprozess) 189.

Gsell B, Krüger W, Lorenz S and Reymann C (ed), Beck-Online Grosskommentar zum BGB (Beck 2023).

Gunning D et al., XAI-Explainable artificial intelligence, 4(37) Science Robotics (2019), https://www.science.org/doi/10.1126/scirobotics.aay7120 accessed 31 December 2023.

Haber E, ‘Privatization of the Judiciary’ (2016) 40(1) Seattle University Law Review 115.

Halis Kasap G, ‘Can Artificial Intelligence (“AI”) Replace Human Arbitrators?’ (2021) (2) Journal of Dispute Resolution 209.

Hartung M, ‘Sonstige Akteure und Rahmenbedingungen’ in M Hartung, M-M Bues and G Halbleib (ed), Legal Tech: Die Digitalisierung des Rechtsmarkts (Beck 2018) 215.

Heetkamp S J and Piroutek C, ‘ChatGPT in Mediation und Schlichtung‘ (2023) 26(3) ZKM (Zeitschrift für Konflikt-management) 80.

Heiss S, ‘Artificial Intelligence Meets European Union Law’ (2021) 10(6) EuCML (Journal of European Consumer and Market Law) 252.

Heldt A, Intensivere Drittwirkung (Mohr Siebeck 2023).

Helfer L and Land M K, ‘The Meta Oversight Board's Human Rights Future’ (2023) 44(106) Cardozo Law Review 2233.

Helmond A, ‘The Platformization of the Web: Making Web Data Platform Ready’ (2015) Social Media and Society 1.

Hess B, Europäisches Zivilprozessrecht (2nd edn, De Gruyter 2021).

Hess B, ‘Prozessuale Mindestgarantien in der Verbraucherschlichtung’ (2015) 70(11) JZ (Juristenzeitung) 548.

Hess T and Waltermann H, ‘Upload-Filter für Content’ (2019) 16(2) MedienWirtschaft 19.

Hoeren T, ‘Sperrpflichten eines Hosting-Anbieters bei rechtswidrigen Informationen sowie wort- und sinngleichen Inhalten’ (2020) (4) LMK (Leitsätze mit Kommentierung) 425949.

Hofmann F, ‘Die neuen Transparenzvorgaben im UWG 2022 im Kontext lauterkeitsrechtlicher Plattformregulierung’ (2022) 124(11) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 780.

Hofmann F‚ ‘Fünfzehn Thesen zur Plattformhaftung nach Art 17 DSM-RL’ (2019) 121(12) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 1219.

Hofmann F, ‘Mittelbare Verantwortlichkeit im Internet’ (2017) 57(8) JuS (Juristische Schulung) 713.

Hofmann F, ‘Prozeduralisierung der Haftungsvoraussetzungen im Medienrecht – Vorbild für die Intermediärshaftung’ (2017) 61(2) ZUM (Zeitschrift für Urheber- und Medienrecht) 102.

Hofmann F and Raue B (ed), Digital Services Act (Nomos 2023).

F Hofmann and L Specht-Riemenschneider, ‘Verantwortung von Online-Plattformen (Responsibility of Online Platforms)’ (2021) 13(1) ZGE (Zeitschrift für geistiges Eigentum) 48.

Hofmann F and Sprenger T, ‘Privatization of Enforcement’ (2021) 85(2) UFITA (Archiv für Medienrecht und Medienwissenschaft) 249.

Holznagel D, ‘Melde- und Abhilfeverfahren zur Beanstandung rechtswidrig gehosteter Inhalte nach europäischem und deutschem Recht im Vergleich zu gesetzlich geregelten notice and take-down-Verfahren‘ (2014) 63(2) GRUR Int. (Gewerblicher Rechtsschutz und Urheberrecht Internationaler Teil) 105.

Holznagel D, Notice and Takedown as Part of Provider Liability (Mohr Siebeck 2013).

Holznagel D, ‘Nutzerrechte bei Facebook: Klärung durch den BGH und bevorstehende Irrwege des EU-Gesetzgebers’ (2021) 37(11) CR (Computer und Recht) 733.

Holznagel D, ‘Zu starke Nutzerrechte in Art. 17 und 18 DSA’ (2022) 38(9) CR (Computer und Recht) 594.

Horwitz J, ‘Facebook Says Its Rules Apply to All. Company Documents Reveal a Secret Elite That’s Exempt’, (2021) Wall Street Journal, https://www.wsj.com/articles/facebook-files-xcheck-zuckerberg-elite-rules-11631541353 accessed 31 December 2023.

Hörnle J, Internet Jurisdiction (Oxford University Press 2021).

Jacques S, Garstka K, Hviid M and Street J, ‘An empirical study of the use of automated anti-piracy systems and their consequences for cultural diversity’ (2018) 15(2) Script-Ed 277.

Janal R, ‘Haftung und Verantwortung im Entwurf des Digital Services Acts’ (2021) 29(2) ZEuP (Zeitschrift für Europäisches Privatrecht) 227.

Jensen O, Tribunal secretaries in international arbitration (Oxford University Press 2019).

Jus Mundi, ‘AI-Powered Search for International Law and Arbitration’, https://jusmundi.com/en accessed 31 December 2023.

Jus Mundi, ‘Jus Mundi Introduces Jus AI: A Game-Changing GPT-Powered AI Solution for the Arbitration Community’ https://dailyjus.com/news/2023/06/jus-mundi-introduces-jus-ai-a-game-changing-gpt-powered-ai-solution-for-the-arbitration-community accessed 31 December 2023.

Kadri T and Klonick K, ‘Facebook v. Sullivan: Public Figures and Newsworthiness in Online Speech Online Speech’ (2019) 93(1) Southern California Law Review 37.

Kaesling K, ‘Evolution statt Revolution der Plattformregulierung‘ (2021) 65(3) ZUM (Zeitschrift für Urheber- und Medienrecht) 177.

Kaesling K, ‘Privatising Law Enforcement in Social Networks: A Comparative Model Analysis’ (2018) 11(3) Erasmus Law Review 151.

Kahneman D, Thinking, Fast and Slow (Farrar, Straus and Giroux 2013).

Kalbhenn J-C, ‘Design Specifications for Chatbots, Deepfakes, and Emotion Recognition Systems’ (2021) 65(8/9) ZUM (Zeitschrift für Urheber- und Medienrecht) 663.

Kaminski M E, ‘Binary Governance: Lessons from the GDPR’s Approach to Algorithmic Accountability’ (2019) 92(6) S. CAL. L. REV. 1529.

Katsh E and Rabinovich-Einy O, Digital Justice (Oxford University Press 2017).

Katyal S, ‘Private Accountability in the Age of Artificial Intelligence’ (2019) 66(1) UCLA Law Review 55.

Katzenbach C, ‘The “Alghorithmic turn” in platform governance’ (2022) 74(1 supp) Cologne Journal of Sociology and Social Psychology 283.

Kaulartz M, ‘Smart Contract Dispute Resolution’ in M Fries and B-P Paal (ed), Smart Contracts (Mohr Siebeck 2019) 73.

Kaye D, Speech Police (Columbia Global Reports 2019).

Keller D, ‘Internet Platforms: Observations on Speech, Danger, and Money’ (2018) Hoover Inst. Aegis Paper Series No 1807.

Kindt T, ‘Blockchainbasierte dezentrale Streitbeilegungsverfahren und ihr Verhältnis zur Schiedsgerichtsbarkeit’ (2023) 21(5) SchiedsVZ (Zeitschrift für Schiedsverfahren) 241.

Kleros, ‘Arbitration Service’, https://kleros.io/en/ accessed 31 December 2023.

Klonick K, ‘The New Governors: The People, Rules and Process Governing Online Speech’ (2018) 131(6) Harvard Law Review 1598.

Klonick K, ‘Why the History of Content Moderation Matters’ (2018) TECHDIRT, <https://www.techdirt.com/articles/20180129/21074939116/whyhistory-content-moderation-matters.shtml> accessed 31 December 2023.

Kosseff J, ‘Private Computer Searches and the Fourth Amendment’ (2018) 14(2) I/S A Journal of Law and Policy 187.

Koulu R, ‘Proceduralizing control and discretion: Human oversight in artificial intelligence policy’ (2020) 27(6) Maastricht Journal of European and Comparative Law 720.

Kreis F, ‘KI und ADR-Verfahren’ in Kaulartz M and Braegelmann T (ed), Rechtshandbuch Artificial Intelligence und Machine Learning (Beck and Vahlen 2020) 633.

Krenzler M and Remmertz F R (ed), Rechtsdienstleistungsgesetz (3rd edn, Nomos 2023).

Krüger W and Rauscher T (ed), Münchener Kommentar zur ZPO (6th edn, Beck 2022).

Kumar S and Kumar H, ‘Mediation and Artificial Intelligence’ (2021) 4(4) International Journal of Law Management & Humanities 1472.

Kumkar L, ‘Plattform-Recht revisited: Umgang mit den Marktordnungen digitaler Plattformen de lege lata et ferenda‘ (2022) 30(3) ZEuP (Zeitschrift für Europäisches Privatrecht) 530.

Kuczerawy A, ‘The Good Samaritan that wasn’t: voluntary monitoring under the (draft) Digital Services Act’ (2021) Verfassungsblog, https://verfassungsblog.de/good-samaritan-dsa/ accessed 31 December 2023.

Kraul T (ed), Das neue Recht der digitalen Dienste – Digital Services Act (DSA) (Nomos 2023).

Ladeur K-H, ‘Neue Institutionen für den Daten- und Persönlichkeitsschutz im Internet: „Cyber-Courts“ für die Blogosphere‘ (2012) 36(10) DuD (Datenschutz und Datensicherheit) 711.

Ladeur K-H, ‘Schutz vor Verletzung von Persönlichkeitsrechten und “Desinformation“ in sozialen Medien unter Bedingungen der politischen Polarisierung’ <https://verfassungsblog.de/personlichkeitsrecht-soziale-medien/> accessed 31 December 2023.

Länderarbeitsgruppe, ‘Abschlussbericht der Länderarbeitsgruppe, “Legal Tech: Herausforderungen für die Justiz’’’ (2019), https://www.schleswigholstein.de/DE/landesregierung/ministerienbehoerden/II/Minister/Justizministerkonferenz/Downloads/190605_beschluesse/TOPI_11_Abschlussbericht.pdf?blob=publicationFile&v=1 accessed 31 December 2023.

Land M, ‘The Problem of Platform Law: Pluralistic Legal Ordering on Social Media’ in P SA Berman (ed), The Oxford Handbook of Global Legal Pluralism (Oxford University Press 2020) 974.

Laukemann B, ‘Private law enforcement and intellectual property: Regulatory challenges in a digital era’ in B Hess, E Jayme and H-P Mansel (ed), Europa als Rechts- und Lebensraum: Liber Amicorum für Christian Kohler zum 75. Geburtstag (Gieseking 2018) 357.

Laukemann B, ‘Private Rechtsdurchsetzung zwischen (digitaler Selbsthilfe) und gerichtlichem Rechtsschutz’ (2022) 8(3) ZfPW (Zeitschrift für die gesamte Privatrechtswissenschaft) 357.

Lawlift GmbH, ‘Lawflift’, https://de.lawlift.com/ accessed 31 December 2023.

Ledwich M and Zautsev A, ‘Algorithmic Extremism: Examining YouTube’s Rabbit Hole of Radicalization’, Cornell University (2019), https://arxiv.org/abs/1912.11211 accessed 31 December 2023.

Leeb C-M, Digitalisierung, Legal Technology und Innovation (Duncker & Humboldt 2019).

Legal Grade AI, ‘Luminace’, https://www.luminance.com/overview.html accessed 31 December 2023.

Leenheer Zimmerman D, ‘A Tale of Legislative Abdication’ (2014) 35(1) Pace Law Review 260.

Leerssen P, ‘An end to shadow banning? Transparency rights in the Digital Services Act between content moderation and curation’ (2023) 48 Computer Law & Security Review 6.

Lennartz J and Kraetzig V, ‘Filtering fundamental Rights’ (2022) Verfassungsblog, https://verfassungsblog.de/filtering-fundamental-rights/ accessed 31 December 2023.

Lessig L, Code and Other Laws of Cyberspace (Basic Books 1999).

Lester T and Pachamanova D, ‘The Dilemma of False Positives: Making Content ID Algorithms More Conducive to Fostering Innovative Fair Use in Music Creation’ (2017) 24(1) UCLA Entertainment Law Review 51.

Lev-Aretz Y, ‘Second Level Agreements’ (2012) 45(1) AKRON L. REV. 137.

Lew J, Mistelis L and Kröll S, Comparative International Commercial Arbitration (Wolters Kluwer 2003).

Lex Machina, ‘Lex Machina Legal Analytics’, https://lexmachina.com/ accessed 31 December 2023.

Liesching M (ed), Netzwerkdurchsetzungsgesetz (Nomos 2018).

Lindquist D and Dautaj Y, ‘AI in International Arbitration’ (2021) (1) Journal of Dispute Resolution 39.

Loo R Van, ‘Federal Rules of Platform Procedure’ (2021) 88(4) University of Chicago Law Review 829.

Loo R Van, ‘The Corporation as Courthouse’ (2016) 33(2) Yale Journal on Regulation 547.

Luhmann N, Legtimation durch Verfahren (11th edn, Suhrkamp 2019).

Lüdemann J, ‘Privatisierung der Rechtsdurchsetzung in sozialen Netzwerken? ’ in M Eifert and T Gostomzyk (ed), Netzwerkrecht (2018) 165.

Lüdemann J, ‘Warum und wie reguliert man digitale Informationsintermediäre?’ in J Lüdemann and Y Hermstrüwer (ed), Der Schutz der Meinungsbildung im digitalen Zeitalter (Mohr Siebeck 2021) 15.

Maier H, Remixes on Hosting Platforms (Mohr Siebeck 2018).

Marmont S, ‘Keeping Up with Legal Technology’ (2019) 1(2) ITA in Review 37.

Mast T, ‘AGB-Recht als Regulierungsrecht’ (2023) 78(7) JZ (Juristenzeitung) 287.

Maxwell G and Vannieuwenhuyse G, ‘Robots Replacing Arbitrators: Smart Contract Arbitration’ (2018) (1) ICC Dispute Resolution Bulletin 24.

Mayer M, Soziale Netzwerke im Internet im Lichte des Vertragsrechts (Richard Boorberg 2018).

McColgan P, ‘Das wird man wohl noch löschen dürfen? – Control Standards for Opinion Rules on the Internet‘ (2021) 1(12) RDi (Recht Digital) 605.

Meder S, ‘Die Zukunft der juristischen Methode: Rehabilitierung durch Chat-GPT?‘ (2023) 78(23) JZ (Juristenzeitung) 1041.

Medvedeva M, Vols M and Wieling M, ‘Using machine learning to predict decisions of the European Court of Human Rights’ (2020) 28(2) Artificial Intelligence and Law 237.

Mendelsohn J K, ‘Die “normative Macht“ der Plattformen – Gegenstand der zukünftigen Digitalregulierung’ (2021) 24(11) MMR (Multimedia und Recht) 857.

Menkel-Meadow C, ‘Is ODR ADR? Reflections of an ADR Founder from 15th ODR Conference, the Hague, the Netherlands, 22-23 May 2016’ (2016) 3(1) IJODR (International Journal on Online Dispute Resolution) 4.

Meta, ‘Community Standards Enforcement Report: Child Endangerment: Nudity and Physical Abuse and Child Sexual Exploitation’, https://transparency.fb.com/data/community-standards-enforcement/child-nudity-and-sexual-exploitation/facebook/ accessed December 2023.

Meta, ‘Community Standards Enforcement Report’, https://transparency.fb.com/reports/community-standards-enforcement/ accessed 31 December 2023.

Meta, ‘How Facebook uses super-efficient AI models to detect hate speech, 19 November 2020’, https://ai.facebook.com/blog/how-facebook-uses-super-efficient-ai-models-to-detect-hate-speech/ accessed 31 December 2023.

Meta, ‘How we review content’, 11 August 2020, https://about.fb.com/news/2020/08/how-we-review-content/ accessed 31 December 2023.

Meta, ‘How we review Content – Prioritization’, 11 August 2020, https://about.fb.com/news/2020/08/how-we-review-content/ accessed 31 December 2023.

Meta, ‘Sharing More Details on How We Will Implement the Oversight Board's Decisions, Responding to the Oversight Board’s First Decisions’, 28 January 2021, https://about.fb.com/news/2021/01/responding-to-the-oversight-boards-first-decisions/ accessed 31 December 2023.

Meta, ‘Facebook Transparency Report of January 2023’, <327151920_907084790305794_6193992151844220602_n.pdf> accessed 31 December 2023.

Meta Transparency Center, ‘Global restrictions’, https://transparency.fb.com/data/content-restrictions/ accessed 31 December 2023.

Meta Transparency Center, ‘How technology detects violations’, 18 October 2023, https://transparency.meta.com/de-de/enforcement/detecting-violations/technology-detects-violations/ accessed 31 December 2023.

Meta Transparency Center, ‘How we assess reports of content violating local law’, https://transparency.fb.com/data/content-restrictions/ accessed 31 December 2023.

Meta Transparency Center, ‘Reviewing high-impact content accurately via our cross-check system’, 12 May 2023, https://transparency.fb.com/enforcement/detecting-violations/reviewing-high-visibility-content-accurately/ accessed 31 December 2023.

Metzger A and Senftleben M, ‘Selected Aspects of Implementing Article 17 of the Directive on Copyright in the Digital Single Market into National Law – Comment of the European Copyright Society’ (2020) 11(2) Journal of Intellectual Property, Information Technology and E-Commerce Law 1.

Milano D, Content control: Digital Watermarking and Fingerprinting (Rhozet 2013).

Monroy M, ‘EU-Kommission droht mit “gesetzgeberischen Maßnahmen“ zur Entfernung von Internetinhalten’, https://netzpolitik.org/2018/eu-kommission-droht-mit-gesetzgeberischen-massnahmen-zur-entfernung-von-internetinhalten/ accessed 31 December 2023.

Monroy M, ‘EU-Internetforum: Viele Inhalte zu “Extremismus“ werden mit Künstlicher Intelligenz aufgespürt’, https://netzpolitik.org/2018/eu-kommission-droht-mit-gesetzgeberischen-massnahmen-zur-entfernung-von-internetinhalten/ accessed 31 December 2023.

Mostert F, ‘Free Speech and Internet Regulation’ (2019) 14(8) Journal of Intellectual Property Law & Practice 607.

Müller-Terpitz R and Köhler M (ed), Digital Services Act (Beck 2024).

Muller C, ‘Setting the Record Straight’, https://blog.youtube/news-and-events/setting-record-straight/ accessed 31 December 2023.

Murgia M, Warell H and Bond D, ‘YouTube revenues under threat over ads alongside extremist videos’ (2017) Financial Times, https://www.ft.com/content/04f8bf56-0b12-11e7-97d1-5e720a26771b accessed 31 December 2023.

Musa S and Bendett S, ‘Islamic Radicalization in the United States – New Trends and a Proposed Methodology for Disruption’ (2010) Washington D.C.: National Defense University, Washington DC Center for Technology and National Security Policy, https://apps.dtic.mil/sti/pdfs/ADA532696.pdf accessed 31 December 2023.

Nahmias Y and Perel M, ‘The oversight of content moderation by AI: Impact assessment and their limitations’ (2021) 58(1) Harvard Journal on Legislation 145.

Nathenson I S, ‘The Procedural Foundations of Information Regulation’ (2020) 24(1) Lewis & Clark Law Review 109.

Nink D, Justiz und Algorithmen (Duncker & Humboldt 2021).

Nolte G, ‘Three Theses on the Current Debate on Liability and Distributive Justice in Hosting Services with User-Generated Content (the so-called “Value Gap” Debate)’ (2017) 61(4) ZUM (Zeitschrift für Urheber- und Medienrecht) 304.

Nunziato D C, ‘The Beginning of the End of Internet Freedom’ (2014) 45(2) Georgetown Journal of International Law 383.

Ohly A and Sosnitza O (ed), Gesetz Gegen den unlauteren Wettbewerb: UWG (8th edn, Beck 2023).

Orssich I, ‘Das europäische Konzept für vertrauenswürdige Künstliche Intelligenz’ (2022) 33(6) EuZW (Europäische Zeitschrift für Wirtschaftsrecht) 254.

Otter, ‘Otter AI’, https://otter.ai accessed 31 December 2023.

Paisley K and Sussman E, ‘Artificial Intelligence Challenges and Opportunities for International Arbitration’ (2018) 11(1) New York Dispute Resolution Lawyer 35.

Pasquale F, The Black Box Society (Harvard University Press 2015).

Peifer K-N, ‘Die neuen Transparenzregeln im UWG (Bewertungen, Rankings und Influencer)’ (2021) 123(12) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 1453.

Perel M and Elkin-Koren N, ‘Accountability in Algorithmic Copyright Enforcement’ (2016) 19(3) Stan. Tech. L. Rev. 473.

Plantin J-C, Lagoze C, Edwards P N and Sandvig C, ‘Infrastructure Studies Meet Platform Studies in the Age of Google and Face-book’ (2016) 20(1) New MEDIA & Society 293.

Polkinghorne M, ‘Different Strokes for Different Folks?’, <https://arbitrationblog.kluwerarbitration.com/2014/05/17/different-strokes-for-different-folks-the-role-of-the-tribunal-secretary-2/> accessed 31 December 2023.

Prütting H, ‘Das neue Verbraucherstreitbeilegungsgesetz: Was sich ändert – und was bleiben wird’ (2016) (3) AnwBl (Anwaltsblatt) 190.

Prütting H, ‘Die rechtliche Stellung des Schiedsrichters’ (2011) 9(5) SchiedsVZ (Zeitschrift für Schiedsverfahren) 233.

Queen Mary University and White & Case, ‘2021 International Arbitration Survey: Adapting Arbitration to a Changing World’, https://arbitration.qmul.ac.uk/research/2021-international-arbitration-survey/ accessed 31 December 2023.

Rabinovich-Einy O and Katsh E, ‘The New New Courts’ (2017) 67(1) Am. U. L. Rev. 165.

Radu R, Negotiating Internet Governance (Oxford University Press 2019).

Rajendra J B and Thuraisingam A S, ‘The deployment of artificial intelligence in alternative dispute resolution: the AI augmented arbitrator’ (2022) 31(2) Information & Communications Technology Law 176.

Raue B, ‘Plattformnutzungsverträge im Lichte der gesteigerten Grundrechtsbindung marktstarker sozialer Netze‘ (2022) 75(2) NJW (Neue Juristische Wochenschrift) 209.

Raue B and Heesen H, ‘Der Digital Services Act‘ (2022) 75(49) NJW (Neue Juristische Wochenschrift) 3537.

Reuter M, ‘Facebook Knew What All Was Going Wrong’ (2021) Netzpolitik.org, https://netzpolitik.org/2021/facebook-files-facebook-wusste-was-alles-schieflaeuft/ accessed 31 December 2023.

Riesenhuber K, ‘§ 10 Die Auslegung’ in K Riesenhuber (ed), Europäische Methodenlehre (De Gruyter 2021) 285.

Richter P and Mendelsohn J, ‘§ 21 Plattformspezifische Vorgaben des Data Acts’ in B Steinrötter (ed), Europäische Plattformregulierung (Nomos 2023) 564.

Rhim Y and Park K, ‘The Artificial Intelligence in International Law’ in E Y J Lee (ed), Revolutionary Approach to international Law: The Role of international Lawyer in Asia (Springer 2023) 215.

Röthemeyer P, ‘Die Schlichtung‘ (2013) 16(2) ZKM (Zeitschrift für Konfliktmanagement) 47.

Ruger T, Kim P, Martin A and Quinn K, ‘The Supreme Court Forecasting Project’ (2004) 104(4) Colum. L. Rev. 1150.

Rule C, ‘Making Peace on eBay’ (2008) ACR Resolution 8.

Rule C, ‘Quantifying the Economic Benefits of Effective Redress: Large ECommerce Data Sets and the Cost-Benefit Case for Investing in Dispute Resolution’ (2012) 34(4) U. ARK. LITTLE ROCK L. REV. 767.

Rühl G, ‘Die Richtlinie über alternative Streitbeilegung: Handlungsperspektiven und Handlungsoptionen’ (2014) 127(1) ZZP (Zeitschrift für Zivilprozess) 61.

Rühl G, ‘Digitale Justiz‘ (2020) 75(17) JZ (Juristenzeitung) 809.

Rühl G, ‘KI in der gerichtlichen Streitbeilegung’ in M Kaulartz and T Braegelmann (ed), Rechtshandbuch Artificial Intelligence und Machine Learning (Beck 2020) 617.

Salter S and Thompson D, ‘Public-Centred Civil Justice Redesign’ (2016-2017) 3 McGill Journal of Dispute Resolution 113.

Saurwein F, ‘Regulierung von Internet-Inhalten: Ombudsstellen als Governance-Option an der Schnittstelle von Recht und Ethik’ in G Marci-Boehncke, M Rath, M Delere and H Höfer (ed), Medien – Demokratie – Bildung (Springer 2022) 47 https://doi.org/10.1007/978-3-658-36446-5_5 accessed 31 December 2023.

Schillmöller J and Doseva S, ‘”Chilling effects” durch YouTubes Content ID?’ (2022) 25(3) MMR (Multimedia und Recht) 181.

Schneiders P, ‘Hate Speech auf Online-Plattformen: Problematization, Regulation and Evaluation against the Background of the Proposal for a Digital Services Act’ (2021) 85(2) UFITA (Archiv für Medienrecht und Medienwissenschaft), https://doi.org/10.5771/2568-9185-2021-2-269 accessed 31 December 2023 accessed 31 December 2023.

Scherer M, ‘Artificial Intelligence and Legal Decision-Making’ (2019) 36(5) Journal of International Arbitration 539.

Scherer M, ‘International Arbitration 3.0. How Artificial Intelligence Will Change Dispute Resolution’ in C Klausegger et al (ed), Austrian Yearbook on International Arbitration (Beck 2019) 503.

Scherer M and Jensen O, ‘Die Digitalisierung der Schiedsgerichtsbarkeit’ in T Riehm and S Dörr (ed), Digitalisierung und Zivilverfahren (De Gruyter 2023), 591.

Schwartz J, ‘Artificial Arbitration?’ in R Wilhelmi and M Stürner (ed), Mehrparteien-Schiedsverfahren: Unter besonderer Berücksichtigung gesellschaftsrechtlicher Streitigkeiten (Springer 2021) 95.

Seetharaman D, Horwitz J and Scheck J, ‘Facebook Says AI Can Enforce Its Rules, but the Company’s Own Engineers Are Doubtful’ (2021) Wall Street Journal, https://www.wsj.com/articles/facebook-ai-enforce-rules-engineers-doubtful-artificial-intelligence-11634338184 accessed 31 December 2023.

Sela A, ‘The Effect of Online Technologies on Dispute Resolution System Design’ (2017) 21(3) Lewis & Clark Law Review 633.

Senftleben M, ‘Institutionalized Algorithmic Enforcement – The Pros and Cons of the EU Approach to UGC Platform Liability’ (2020) 14(2) FIU Law Review 299.

Shaughnessy P and Rogers C, ‘Arbitrator Intelligence - An Interview with its Founder and Director, Professor Catherine Rogers’ (2015) 1(1) Journal on Technology in International Arbitration 87.

Shinn L D, ‘YouTube’s Content ID as a Case Study of Private Copyright Enforcement Systems’ (2015) 43(2/3) AIPLA Quarterly Journal 359.

Silicon Valley Arbitration and Mediation Center, ‘Silicon Valley Arbitration & Mediation Center Guidelines, Draft of August 31, 2023’, https://thearbitration.org/wp-content/uploads/2023/08/SVAMC-AI-Guidelines-CONSULTATION-DRAFT-31-August-2023-1.pdf accessed 31 December 2023.

Sim C, ‘Will Artificial Intelligence Take Over Arbitration?’ (2018) 14(1) Asian International Arbitration Journal 1.

Simshaw D, ‘Access to A.I. Justice: Avoiding an Inequitable Two-Tiered System of Legal Services’ (2022) 24 Yale Journal of Law & Technology 150.

Snijders H, ‘Arbitration and AI, Arbitration 2023’ in H Snijders (ed), Arbitration and AI, Arbitration (Wolters Kluwer 2023) 224.

Solomon L, ‘Fair users or content abusers’ (2015) 44(1) Hofstra L. Rev. 237.

Specht F, ‘Chancen und Risiken einer digitalen Justiz für den Zivilprozess’ (2019) 22(3) MMR (Multimedia und Recht) 153.

Spindler G, ‘Der Vorschlag für ein neues Haftungsregime für Internetprovider – der EU-Digital Services Act (Teil 1)’ (2021) 123(4) GRUR (Gewerblicher Rechtsschutz und Urheberrecht) 545.

Spoerri T, ‘On Upload Filters and other Competitive Advantages for Big Tech Companies under Article 17 of the Directive on Copyright in the Digital Single Market’ (2019) 10(2) Journal of Intellectual Property, Information Technology and Electronic Commerce Law 173.

Staudinger J von von (ed), Kommentar zum BGB, Buch 2 (19th rev edn, De Gruyter 2022).

Stotz R, ‘Die Rechtsprechung des EuGH’ in K Riesenhuber (ed) Europäische Methodenlehre (De Gruyter 2021) 653.

Sunstein C R, Republic – Divided Democracy in the Age of Social Media (Princeton University Press 2017).

Suzor N P, Lawless: The Secret Rules that Govern our Digital Lives (Cambridge University Press 2019).

Suzor N P and Gray J E, ‘Playing with machines: Using machine learning to understand automated copyright enforcement at scale’ (2020) 7(1) Big Data & Society, https://doi.org/10.1177/2053951720919963 accessed 31 December 2023.

Sweeney L, ‘Discrimination in Online Ad Delivery’ May 2013 Comm. ACM 44, http://cacm.acm.org/magazines/2013/5/163753-discrimination-in-online-ad-delivery/ accessed 31 December 2023.

Taeger J and Kremer S, Recht im E-Commerce und Internet (Beck 2021).

Tan V, ‘Online Dispute Resolution for Small Civil Claims in Victoria’ (2019) 24 Deakin Law Review 101.

Titlow J P, ‘Youtube is using AI to police copyright to the tune of $2 billion in payouts’, 31 July 2016, https://www.fastcompany.com/4013603/youtube-is-using-ai-to-police-copyright-to-the-tune-of-2-billion-in-payouts accessed 31 December 2023.

Trint, ‘Trint’, https://trint.com/ accessed 31 December 2023.

Tushnet R, ‘All of this has happened before and all of this will happen again: Innovation in copyright licensing’ (2014) 29(3) Berkeley Technology Law Journal 1447.

Twitter, ‘Network Enforcement Report: January-June 2023’, https://transparency.twitter.com/content/dam/transparency-twitter/country-reports/germany/NetzDG-Jan-Jun-2023.pdf accessed 31 December 2023.

Tyler T R, ‘What is procedural justice?: Criteria used by citizens to assess the fairness of legal procedures’ (1988) 22(1) Law & Society Review 103.

Tyler T R, Why people obey the law (Princeton University Press 2006).

Urban J, Karagani J and Schofield B, ‘Notice and Takedown in Everyday Practice’ (2016) UC Berkeley Public Law Research Paper No 2755628, https://ssrn.com/abstract=2755628 or http://dx.doi.org/10.2139/ssrn.2755628.

Valkanova M, ‘Trainieren von KI-Modellen’ in M Kaulartz and T Braegelmann (ed), Rechtshandbuch Artificial Intelligence und Machine Learning (Beck 2020) 336.

Vannieuwenhuyse G, ‘Arbitration and New Technologies: Mutual Benefits’ (2018) 35(1) Journal of International Arbitration 119.

Voß W, ‘Gerichtsverbundene Online-Streitbeilegung‘ (2020) 84(1) RabelsZ (Rabels Zeitschrift für ausländisches und internationales Privatrecht) 62.

Wagner B, Global Free Expression – Governing the Boundaries of Internet Content (Springer 2016).

Wagner G, ‘Haftung von Plattformen für Rechtsverletzungen (Teil 2)’ (2020) 122(5) GRUR (Gewerblicher Rechts-schutz und Urheberrecht) 447.

Wagner G, ‘Private Law Enforcement and ADR’ in J Zekoll, M Bälz and I Amelung (ed), Formalisation and Flexibilisation in Dispute Resolution (Brill 2014) 369.

Wagner J, Legal Tech und Legal Robots (2nd edn, Springer Gabler 2020).

Walker K, ‘Four ways Google will help to tackle extremism’ (18 June 2017) Financial Times, https://www.ft.com/content/ac7ef18c-52bb-11e7-a1f2-db19572361bb accessed 31 December 2023.

Wall Street Journal, ‘The Facebook Files’ (13 September 2021), https://www.wsj.com/articles/the-facebook-files-11631713039 accessed 31 December 2023.

Welser M von, ‘Die KI-Verordnung – ein Überblick über das weltweit erste Regelwerk für künstliche Intelligenz’ (2024) 16(15) GRUR-Prax (Gewerblicher Rechtsschutz und Urheberrecht in der Praxis) 485.

Wendland M, Mediation und Zivilprozess (Mohr Siebeck 2017).

Wielsch D, ‘Die Ordnungen der Netzwerke, AGB – Code – Community Standards‘ in M Eifert and T Gostomzyk (ed), Netzwerkrecht (Nomos 2018) 61.

Wielsch D, ‘Medienregulierung durch Persönlichkeits- und Datenschutzrechte‘ (2020) 75(3) JZ (Juristenzeitung) 105.

Wilske S, Markert L and Ebert B, ‘Entwicklungen in der internationalen Schiedsgerichtsbarkeit im Jahr 2022 und Ausblick auf 2023’ (2023) 21(3) SchiedsVZ (Zeitschrift für Schiedsverfahren) 121.

Wolfowicz M, Weisburd D and Hasisi B, Examining the interactive effects of the filter bubble and the echo chamber on radicalization, (2023) 19(5) Journal of Experimental Criminology 119.

Wolters Kluwer, ‘Arbitration Database’, https://www.kluwerarbitration.com/ accessed 31 December 2023.

Wolters Kluwer, ‘smartlaw’, https://www.smartlaw.de/ accessed 31 December 2023.

Wu T, ‘Will Artificial Intelligence Eat the Law? The Rise of Hybrid Social-Ordering Systems’ (2019) 119(7) Columbia Law Review 2001.

Yeung K, ‘”Hypernudge”: Big Data as a mode of regulation by design, Information’ (2017) 20(1) Communication & Society 118.

Yablon Y and Landsman-Ross N, ‘Predictive Coding’ (2013) 64(3) South Carolina Law Review 633.

YouTube, ‘Accelerated Content ID & Complaint Process’, https://support.google.com/youtube/thread/171619847 accessed 31 December 2023.

YouTube, ‘Answers to common questions about Copyright claims on YouTube’, https://support.google.com/youtube/thread/1281991 accessed 31 December 2023.

YouTube, ‘Appeal a Content ID claim’, https://support.google.com/youtube/answer/12104471 accessed 31 December 2023.

YouTube, ‘Best practices for claims’, https://support.google.com/youtube/answer/4352063 accessed 31 December 2023.

YouTube, ‘Best practices for references’, <https://support.google.com/youtube/answer/107008> accessed: 31 December 2023.

YouTube, ‘Content eligible for Content ID’, https://support.google.com/youtube/answer/2605065?hl=en accessed 31 December 2023.

Youtube ‘Copyright Strike Basics’ https://support.google.com/youtube/answer/2814000#zippy=%2Cfolgen-einer-urheberrechtsverwarnung accessed 31 December 2023.

YouTube, ‘Copyright Transparency Report H2 2021’, https://storage.googleapis.com/transparencyreport/report-downloads/pdf-report-22_2021-7-1_2021-12-31_en_v1.pdf accessed 31 December 2023.

YouTube, ‘Create an asset’, https://support.google.com/youtube/answer/3011552?hl=en&ref_topic=3011550 accessed 31 December 2023.

YouTube, ‘Deliver content using spreadsheet templates’, https://support.google.com/youtube/answer/6066171 accessed 31 December 2023.

YouTube, ‘Dispute a Content ID claim’, <https://support.google.com/youtube/answer/2797454?hl=en&ref_topic=9282678#zippy=%2Coptionen-f%C3%BCr-den-anspruchsteller> accessed 31 December 2023.

YouTube, ‘Fix reference overlaps’, https://support.google.com/youtube/answer/3022604?hl=en&ref_topic=3013248 accessed 31 December 2023.

YouTube, ‘Frequently asked questions about copyright’, https://support.google.com/youtube/answer/2797449?hl=en accessed 31 December 2023.

YouTube, ‘Frequently asked questions about fair use’, https://support.google.com/youtube/answer/6396261#zippy=%2Ci-posted-a-disclaimer-on-my-video%2Ci-gave-credit-to-the-copyright-owner%2Cim-using-the-content-for-entertainment-or-non-profit-uses%2Cwhen-does-fair-use-apply%2Cwhat-constitutes-fair-use%2Chow-does-fair-use-work%2Chow-does-content-id-work-with-fair-use accessed 31 December 2023.

YouTube, ‘How policies are applied’, https://support.google.com/youtube/answer/3369929 accessed 31 December 2023.

YouTube, ‘Monetization during Content ID disputes’, https://support.google.com/youtube/answer/7000961?hl=en&ref_topic=9282678 accessed 31 December 2023.

YouTube, ‘Qualify for Content ID’, https://support.google.com/youtube/answer/1311402 accessed 31 December 2023.

YouTube, ‘Requirements for counter notifications’, https://support.google.com/youtube/answer/6005919?hl=en&ref_topic=9282678 accessed 31 December 2023.

YouTube, ‘Review potentially invalid references’, https://support.google.com/youtube/answer/6013183 accessed 31 December 2023.

YouTube, ‘Update: Improving Content ID for creators’, https://blog.youtube/news-and-events/update-improving-content-id-for-creators/ accessed 31 December 2023.

YouTube, ‘Use Content ID matching on live streams’, https://support.google.com/youtube/answer/9896248?hl=en accessed 31 December 2023.

YouTube, ‘Using Content ID’, https://support.google.com/youtube/answer/3244015?hl=en accessed 31 December 2023.

YouTube, ‘Using the YouTube DDEX feed’, https://support.google.com/youtube/topic/3505247 accessed 31 December 2023.

YouTube, ‘What are policies?’, https://support.google.com/youtube/answer/107383?hl=en&ref_topic=24332 accessed 31 December 2023.

YouTube, ‘What Does Fair Use Mean’, https://support.google.com/youtube/answer/9783148?hl=de accessed 31 December 2023.

YouTube, ‘What is copyright?’, https://support.google.com/youtube/answer/2797466#refrained&zippy=%2Cmissverst%C3%A4ndnis-nr-if-you-angive-that-your-content-does-not-serve-commercial-purposes-you-can-use-any-content accessed 31 December 2023.

YouTube, ‘What is a reference?’ https://support.google.com/youtube/answer/107004?hl=en accessed 31 December 2023.

YouTube, ‘Where can I get more information about copyright outside the U.S.?’ https://support.google.com/youtube/answer/2797449?hl=de&ref_topic=2778546#zippy=%2Cwo-erhalte-ich-weitere-informationen-zum-urheberrecht-außerhalb-der-usa accessed 31 December 2023.

YouTube, ‘YouTube Partner Program overview & eligibility’ https://support.google.com/youtube/answer/72851 accessed 31 December 2023.

Zawada K L, ‘The Emergence and Development of Content ID in Light of User-generated Law, in How Deep is your Law?’ (2017) 5th International Conference of PhD Students and Young Researchers Conference Papers 483.

Zekos G, Advanced Artificial Intelligence and Robo-Justice (Springer 2022).

Zhou T, Postmorten: ‘Every Frame a Painting (2017)’ https://perma.cc/U5WU-M6ZZ accessed 31 December 2023.

Zorrilla E, ‘Towards a Credible Future’ (2018) 16(2) SchiedsVZ (Zeitschrift für Schiedsverfahren) 106.

***

---