8

Under Article 14 of the E-Commerce Directive, which remains in force and will be replicated largely unchanged by the Digital Services Act [7], hosting services (which include social media) are immune from liability for making available illegal content posted by users, as long as they are not aware of the illegal content or remove it expeditiously on becoming aware of it. In practice, this has created a notice-and-takedown regime in which aggrieved parties can contact platforms to inform them about illegal content, with the result that the platform must remove it to avoid liability. [8] However, this general immunity is now complicated by three developments.

9

First, Article 14 precludes civil or criminal liability for user-generated content, but not injunctions. Since the E-Commerce Directive’s introduction, injunctive relief has in particular played a key role in copyright enforcement. [9] More recently, the ECJ has accepted the use of injunctions to impose stringent moderation obligations on social media platforms. In its controversial Glawischnig-Piesczek [2019] decision, the ECJ upheld an Austrian court’s imposition of an injunction requiring Facebook not only to delete posts which had been held to defame the claimant, but to find and delete, on an ongoing basis, all identical or equivalent content. [10] This marks a significant shift from its earlier rulings in Scarlet v SABAM [2011] and SABAM v Netlog [2012] that injunctions could not require an internet service provider to actively check all user uploads for copyright-infringing content. [11]

10

In Glawischnig-Piesczek, both the judgment and the Advocate General’s opinion attached significant weight to the supposed availability of technological tools that could automatically detect content equivalent to that deemed illegal. [12] Experts consider this confidence in automated moderation tools unwarranted. They remain highly unreliable [13], and their use poses severe risks to users’ freedom of expression and privacy rights. [14] Nonetheless, given the political pressure on platforms to take action on harmful content and the at-least-apparent promise that AI technologies can enable more comprehensive enforcement, the use of injunctions to impose such sweeping moderation obligations may become more common.

11

Second, the EU has introduced different liability regimes in some areas, specifically for terrorist content (under the 2021 Terrorist Content Regulation [15]) and copyright infringement (under the 2019 Copyright Directive [16]). The Terrorist Content Regulation requires platforms to remove terrorist content (which is broadly and vaguely defined, such that it could frequently include journalistic content [17]) within one hour after receiving a removal order from law enforcement. [18] They may also be required by competent national authorities to take further proactive measures to find and remove terrorist content. [19] Article 17 of the Copyright Directive, on the other hand, creates a new liability regime in which platforms are treated as primarily liable for copyright infringement unless they make best efforts to obtain a license from the rightsholder and, in the absence of a license, make best efforts to remove copyright works which have been notified to them by rightsholders and prevent all future uploads. [20] The latter obligation is widely acknowledged by academic experts [21], and by the Advocate General in his recent opinion in Poland’s judicial review case against Article 17 [22], to require automated filtering of all user uploads in order to identify and block the notified copyright works.

12

Both pieces of legislation were highly controversial, due in large part to the perceived risks of “overblocking” of legal content. [23] Kuczerawy [24] and Frosio and Mendis [25] have suggested that, in combination with the Glawischnig-Piesczek ruling, these laws could mark the abandonment of the foundational principle in Article 15 E-Commerce Directive, as interpreted by the ECJ in the SABAM cases, that platforms cannot be under a general obligation to monitor all content for illegality. The principle has effectively been reinterpreted, such that an impermissible general monitoring obligation will not be taken to exist as long as platforms are only required to search for certain specific content, even if all content on the platform must be monitored for that purpose. [26]

13

Finally, at the same time, some member states have introduced national measures requiring deletion of illegal content within short time limits, such as the German NetzDG [27], Austrian Kommunikationsplattformen-Gesetz  [28], and French loi Avia (although most provisions of the latter were struck down by the Constitutional Council in June 2021 [29]). While these laws can be regarded as simply specifying in more detail what constitutes “expeditious” removal under Article 14 E-Commerce Directive, their compatibility with the Directive is questionable, given that its aim was to create harmonised EU-wide standards and that it calls for platforms to be regulated only in the EU member state where they are headquartered. [30]

14

A second feature of the developing regulatory landscape is the active encouragement of private ordering, through the encouragement of self-regulation and the creation of legal duties outside the intermediary liability framework. [31] Article 5 of the Terrorist Content Regulation requires platforms designated by regulators as exposed to terrorist content to take “specific measures” to address it. These measures remain largely at their own discretion, though one example specified in the provision is introducing new restrictions in their contractual community standards—a notable step towards privatised enforcement. [32] Article 17(10) Copyright Directive and Section 5 of the proposed Digital Services Act both mandate the Commission to work with businesses to develop industry codes and best practices. [33] Such co-regulatory measures have already significantly affected how platforms moderate both legal and illegal content, encouraging them to go beyond notice-and-takedown regimes and introduce more proactive content removal measures, including increasing use of automated moderation. [34]

15

European policymakers have also placed informal pressure on platforms to introduce new content governance measures, often with the threat that harder regulation will otherwise follow. [35] In response to rising public and political concerns about racist hate speech and disinformation following the 2015 “refugee crisis”, the 2016 Brexit referendum and the 2016 US election, leading European policymakers initially showed a clear preference for encouraging industry self-regulation. [36] The Commission negotiated a Code of Conduct on Hate Speech and Code of Practice on Disinformation with leading platforms in 2016 and 2018 respectively. [37] Informal pressure from European policymakers was also instrumental in leading major platforms to set up the GIFCT, an industry body which coordinates the removal of terrorist content across all participating platforms, using a hash database to flag any future uploads that are identical to previously removed content. [38]

18

Schultz argues that the sexual model of workplace sex discrimination both ignores and distracts from other important aspects of discrimination: it diverts employers’, employees’ and the courts’ attention from sexist conduct which is not sexual in nature and from structural discrimination which cannot be reduced to individual misconduct. In platform regulation, it is important to question whether liability for certain types of illegal content distracts attention from other issues. Liability risks evidently influence how platforms allocate resources to moderation and other “trust and safety” programmes: this is illustrated by the major platforms’ immediate investment of significant additional resources and moderation staff in Germany following the introduction of NetzDG. [41] However, as recent leaks from within Facebook revealed, even the biggest and wealthiest tech companies make very limited resources available for trust and safety projects. [42] Any deployment of resources and personnel to areas that do not generate revenue is unlikely to be approved by private corporations unless there is another clear financial justification, such as regulatory compliance. Thus, it can be assumed that any regulation requiring platforms to invest resources in one aspect of content governance risks reducing the resources available to investigate and address other social issues.

19

Like the narrow definition of sex discrimination that Schultz criticises, the tendency in European regulation to single out illegal content for deletion risks diverting attention from other types of harmful behaviour. Taking hate speech as an example, Ben-David and Matamoros-Fernández have documented how hate can systematically be spread on social media through content that does not itself fall under hate speech bans. [43] For example, users can post something just within the law which encourages hate speech in the comments, or like and comment on posts containing hate speech to increase their visibility to other users. Focusing only on the legality of content (posts, comments etc.) also ignores other types of abusive behaviour, such as coordinated malicious reporting of other users for legal or policy violations. [44] This may even be actively facilitated by rules requiring expeditious removal of illegal content, since incentivising quick responses may increase the likelihood of spurious complaints being upheld. There is anecdotal evidence of coordinated malicious reporting being used against victims of discrimination under the German NetzDG system. [45]

20

Moreover, most EU regulation overlooks the structural factors which contribute to policy problems like online racism, disinformation and discrimination. In general, it targets the level of individual pieces of content—not only by requiring illegal content to be removed, but also by providing safeguards for freedom of expression which largely involve individual users complaining that their individual posts should be reinstated. [46] Focusing only on the content level fails to take into account how the harmfulness of content can differ widely depending on its context. [47] For example, one of the most harmful aspects of online harassment is its networked nature: users can easily incite others to join them in harassing a target with large numbers of abusive messages and other harmful actions, such as revealing personal information. [48] In such cases, examining the legality of individual messages may entirely overlook the primary harm they cause, as well as being practically unlikely to address enough of the harassment to have a significant impact.

21

Focusing on the content level also ignores important contextual and structural factors. Even in instances where harm is inflicted by individual pieces of content and can be remedied by content removal, considering contextual factors such as platform design and user cultures is crucial to ensure effective moderation. For example, much harmful behaviour is not reported to moderators because platform interfaces make it laborious for users to report it or because they do not expect a helpful response. [49] More broadly, structural factors such as platform algorithms, architectures and business models can contribute to significant social harms which cannot be resolved by removing individual pieces of content.

22

The typical social media business model, which is based on maximising user engagement and time on site in order to gather as much data and sell as much advertising space as possible, is frequently criticised for exacerbating social harms such as hate speech and disinformation. In particular, algorithms optimised for maximum user engagement are accused of promoting divisive, extremist and sensationalist content, and driving users towards harmful content and ideologies by showing them more extreme versions of whatever they are interested in. [50] Systematic studies of this phenomenon are lacking (and are hampered by the inaccessibility of platform data to independent researchers). [51] However, there is some evidence to support these claims. Journalistic investigations have found that Facebook, Instagram, YouTube and TikTok (all of which rely heavily on algorithmic content ranking and recommendations) actively recommend extremist content, as well as other harmful content such as self-harm, and show increasingly extreme content to users based on their previous interests. [52]

23

Platforms’ profiling and categorisation of users can also have more subtle impacts, such as reinforcing social inequalities. To target content and ads, platforms commonly profile users based on sensitive identity categories like gender and race, often using simplistic and offensive categorisations (e.g. imposing binary gender categories irrespective of user preferences). [53] These tend to symbolically further marginalise historically oppressed groups, by positioning them as deviations from a default “normal” user who is white, straight, etc. [54] They can also materially harm such groups in various ways: for example, by exposing sensitive information to advertisers [55], allowing advertisers to deliberately target vulnerable groups [56], or excluding them from economic opportunities.

24

A particularly well-studied example which obviously replicates historical patterns of discrimination is when marginalised users are excluded from adverts for jobs or housing. Facebook in the past allowed advertisers to explicitly exclude certain “ethnic affinities” from their ad audiences, which attracted heavy criticism. [57] However, researchers have shown that even without using criteria referring to race or other protected characteristics, advertisers can use proxies such as language or place of residence to exclude certain groups. [58] Moreover, even where there is no intention to discriminate, predictive targeting may automatically select audiences which are heavily skewed by race, gender and other protected characteristics [59]: this may, for example, reinforce the disadvantage women face in many professions by preventing them from seeing adverts for jobs that have historically been more appealing to men. [60] The use of predictive “affinity profiling” rather than concrete data about how users identify may allow such profiling to escape the ambit of anti-discrimination and data protection law. [61]

25

The failure to address structural issues such as these, and the near-exclusive focus on illegal content as the key vector for harm, is a major flaw of the current European approach to platform regulation. It should be noted that the Digital Services Act represents a partial shift away from this approach, in that it introduces new obligations for platforms to assess and take action on “systemic risks stemming from the functioning and use of their services”. [62] Article 27 explicitly encourages them to make structural changes in order to mitigate these risks, such as altering platform design and algorithms, or reforming internal processes and organisation. [63] This represents a positive step away from a narrowly content-focused approach.

26

However, these changes should not be overstated. First, the relevant obligations apply only to the category of “very large online platforms”, those with over 45 million EU users. [64] Smaller platforms also have new obligations, but these mostly address the content level (e.g. complaints and redress mechanisms for individual content removal decisions). Second, how effective the new regulations for very large online platforms will be in practice remains to be seen. They rely heavily on self-regulation and privatised enforcement. While the Commission will have new oversight powers including the right to require disclosure of information from very large online platforms and to conduct on-site inspections [65], the primary procedures intended to identify and address systemic risks will be platforms’ internal risk assessments and voluntary measures, and yearly independent expert audits. [66] These types of privatised regulatory enforcement are intransparent and prone to capture [67], especially in complex, high-tech, information-based industries—such as social media—where external oversight is difficult. [68]

27

In another influential critique of Title VII, Edelman theorised a process of “legal endogeneity” whereby formalities used to demonstrate compliance come to eclipse the substantive goals of regulation. [69] This allows businesses to influence the law to their own advantage, as courts and regulators increasingly defer to industry “best practices” when deciding whether legal standards have been met. Edelman’s theory has been applied to technology regulation by Waldman [70], who finds ample evidence for similar processes taking place in privacy law enforcement. The Digital Services Act’s regulatory approach may create similar problems, with formalities like risk assessments taking precedence over meaningful change in industry practices and ultimately reinforcing the status quo. A regulatory focus on mitigating discrete risks also overshadows broader questions about how technologies are used and for whose benefit. [71] Typically, harms that diverge from what powerful industry actors deem “normal” are classified as risks, while harms that stem from underlying structural features of an industry are not. [72] As will be discussed in more detail in section 4, when private actors are charged with the definition and identification of risks, they will tend to construct those risks in the ways that best serve their own business interests.

28

Equally, EU regulation of social media content is overinclusive in significant respects. Like the American sex discrimination jurisprudence that Schultz criticises, it incentivises platforms to delete and suppress a wide range of content and behaviour that should not be considered harmful. Perhaps the best-documented example is the suppression by almost all major platforms of content that is sexually suggestive and/or related to sex work. [73] This causes significant material harm to sex workers by cutting off income sources, driving them towards more dangerous offline work and preventing them from advocating politically for their interests. [74] Blanket bans on sexual content also affect other users’ wellbeing, for example by hampering access to sexual health advice [75], and lead to much broader policing of online art, culture and self-expression. For example, museums have regularly been blocked from posting images of nude art when promoting exhibitions. [76] Such policies could aptly be described as creating a “sanitised” internet.

29

These strict policies are significantly influenced by the US’ 2018 FOSTA/SESTA legislation, which removed platforms’ intermediary liability exemptions for content related to sex work. [77] Many platforms which did not already ban sexual content for commercial reasons responded to the legislation by implementing strict bans on sexual content worldwide, including in countries where sex work is legal. [78] However, the impact of European regulatory choices in this context should not be overlooked. First, if European legal systems did not grant platforms near-unfettered discretion to remove legal content under their contractual terms of service, they would not be able to arbitrarily impose US standards worldwide. Second, an important factor driving platforms to ban sexual content is app store policies: social media platforms rely heavily on users accessing them through mobile apps, and Apple (one of the two dominant app stores) is particularly notorious for banning apps that permit any kind of sexual content. [79] While the Commission is currently investigating Apple’s App Store for anticompetitive behaviour relating in particular to its enforcement of in-app payments from which it takes a commission [80], European authorities have chosen not to intervene in Apple’s use of its infrastructural power to enforce content policies that suppress sexual content across a wide swathe of the internet. Finally, some European countries have similar laws restricting pornography and advertising for sex work: examples include the German Jugendschutzgesetz, which sets broad requirements for online media accessible to under-18s to be child-friendly [81], and Article 380ter of the Belgian Criminal Code, which criminalises all advertising of sex work. [82] These would in any case incentivise platforms to take a restrictive approach.

30

Similar over-enforcement can be seen in regard to other types of content. Over-removal in copyright cases, based on obviously spurious notices from rights-holders, has been extensively documented. [83] Commentators have raised particular concerns about the inability of automated classifiers to identify legally protected uses of a work such as parody and quotation. [84] Copyright notices have also been abused to effect the removal of political content. [85] Attempts by platforms to remove terrorist content regularly censor activists aiming to challenge extremism or document violent incidents. [86] There have also been numerous documented instances of social media posts in which people of colour describe their experiences of racism being tagged as racist hate speech and deleted [87], or reclaimed slurs that are widely used in a positive sense in LGBTQ+ communities being indiscriminately censored. [88]

31

Schultz highlights that the over-enforcement of harassment law is not evenly distributed, but reflects existing inequalities and power structures. She describes cases where sexual harassment claims were used to target LGBTQ+ employees, or where sexualised behaviour which was tolerated from white employees was treated as inappropriate when it came from people of colour. Similarly, the disproportionate impact of online content moderation on minorities and marginalised groups has been well documented. Policies on sexual content and nudity not only frame female and non-binary bodies as problematic [89]; they have also consistently been disproportionately enforced against women of colour and people who do not meet normative beauty standards, while celebrities and conventionally attractive white women are treated more leniently. [90] Waldman has also comprehensively detailed how the suppression of sexual content disproportionately affects LGBTQ+ users, maintaining social media platforms as “straight spaces”. [91] Major platforms often permit queer visibility only where it is desexualised, unthreatening and integrated into heteronormative family structures and values. [92] In the context of terrorist content—a regulatory priority for the EU—moderation unfolds through close cooperation between platforms and European security agencies [93], which primarily target Islamist terrorism and have long histories of racist and Islamophobic discrimination. [94] Bloch-Wehba has shown how the way platforms define and identify terrorist content is heavily shaped by security discourses which have consistently stigmatised and targeted Muslims, while downplaying threats from the extreme right. [95] This appears to be one reason that Arabic social media users—including activists and journalists—are particularly vulnerable to indiscriminate censorship. [96]

33

In the tradition of feminist and queer legal theory, Schultz problematises the supposedly clear legal categories on which the allocation of liability is based. She argues that clearly defining sexuality and walling it off from other aspects of social life is impossible; attempts to do so typically enforce dominant norms around sexual conduct and are imbued with bias against minority groups. The same could be said of defining “terrorist content”, a broad and vague category which has long been used to legitimise anti-Muslim bias [97]; or even of “hate speech”, a category which is meant to protect marginalised groups. Hate speech remains a deeply contested concept, and its interpretation is influenced by established social norms and power structures. As Post highlights, the term is rarely applied to elite discourse, even where it has evident discriminatory effects. [98] In practice, it has been used by social media platforms to suppress marginalised groups’ challenges to oppressive social structures. [99] Overall, when European policymakers exhort platforms to be “responsible” [100] and act in accordance with European values [101], they are strategically glossing over the contested nature of these values.

34

Schultz also makes a forceful case for a sociolegal approach that highlights the gaps between how the “law on the books” allocates liability and how businesses respond to liability incentives in practice. This is closely related to the former point since ambiguous legal categories give businesses greater latitude for selective and self-interested enforcement. In the context of social media regulation, Frosio and Husovec have highlighted how formal legal liabilities are just one factor influencing platforms’ content governance: “The real responsibility landscape is equally determined by a mixture of voluntary agreements, self-regulation, corporate social responsibility, and ad hoc deal-making.” [102] This is especially and increasingly the case as the EU promotes private ordering measures such as self-regulation and flexible legal obligations based on industry “best practices”, as outlined in section 2.1.

35

This has implications for the normative orientation of the law. Edelman and Waldman’s work on legal endogeneity shows empirically how, when the law charges private actors with enforcing vaguely-defined legal standards, they are likely to be interpreted in a way that serves corporate interests and dominant social norms more than the nominal goals of the regulation—even where these are supposedly progressive and egalitarian. [103] Schultz further argues that corporate actors will interpret the law in simplified ways to streamline enforcement processes, over-enforce to minimise liability risks and focus on suppressing economically unproductive behaviour over behaviour which is actually harmful.

36

These problems are equally present in social media regulation. Speech rules must be simplified and streamlined to enable industrial-scale content moderation for global platforms [104]: the injustices that can result from such reductive interpretations are exemplified by the 2017 leak revealing that Facebook’s content moderation guidelines defined invective against white men as “hate speech” but not equivalent content targeting black children. [105] This dynamic is exacerbated by increasing reliance on algorithmic enforcement, which is actively encouraged by EU law, given the limitations of currently-existing technology in understanding the meaning and context of expressions. [106] Speech rules shift to reflect what algorithms are capable of assessing, rather than what is actually considered desirable on policy grounds: for example, when all nudity is treated as pornography because it is what can most easily be identified by image recognition software. [107]

37

Overblocking to minimise liability risks is also a much-discussed problem [108], and the influence of platforms’ economic interests on their content moderation practices is evident. Content moderation experts point out that apparent inconsistencies in moderation policies tend to line up with whether the content in question is valuable for advertisers. [109] Recalling the ambiguities of the term “hate speech” discussed above, it is notable that major social media companies have openly negotiated with the World Federation of Advertisers to align the definition of hate speech in their platform content policies with what advertisers consider harmful to their “brand safety”. [110]

38

Considering how commercial priorities shape the application of the law is particularly important given the increasing turn towards private ordering in EU platform regulation. Platforms are not only being co-opted to enforce state speech regulation. [111] They are required to make “best efforts” on enforcement [112], choose the appropriate “specific measures” to respond to harmful content [113], utilise contractual terms and conditions to forbid harmful behaviour [114], and agree self-regulatory industry codes and best practices. [115] As Land suggests, these very broad discretionary powers over how the law is interpreted and how offline norms are adapted to the online context effectively amount to legislative power. [116]

39

In the context of sexual harassment, Schultz shows that this delegation of power leads to a wide discrepancy between what the law states is illegal and what is actually banned in workplaces in practice. Similarly, delegating the interpretation of speech laws to platforms can significantly change what they are understood to mean. For example, contextual factors that are traditionally considered relevant in applying the law but are harder to incorporate into industrial-scale moderation processes may be excluded entirely. [117] As noted above, this is exacerbated by automated enforcement, as standards shift to reflect the limited evaluative capabilities of software. [118] Safeguards provided by law—such as appeals systems for users, which the EU relies upon heavily in the Terrorist Content Regulation, Copyright Directive and Digital Services Act [119]—may not be effective or widely used in practice. [120] For example, Bloch-Wehba argues that where regulations heavily incentivise automated removal but stipulate that appeals should involve human review, in practice this will mean that the former takes place at scale but the latter cannot. [121]

40

As well as disproportionately affecting marginalised groups through specific enforcement decisions, such private ordering is likely to more broadly reinforce mainstream or dominant norms about permissible views, discourse and sexual expression. Regulators’ appeals for platforms to act “responsibly” and in accordance with public values [122] may risk incentivising a majoritarian approach, where platforms simply try to regulate content in line with dominant tastes and ideologies, while suppressing controversial or non-mainstream viewpoints—as observed by Waldman in his study of platforms as “straight spaces”. [123]

41

Moreover, the EU’s reliance on private ordering measures means that enforcement of regulatory objectives is in practice inseparably intertwined with platforms’ pursuit of their own commercial goals. As discussed in section 2.2, Tom Nichini TN @lydia cross-reference platforms are encouraged by the Copyright Directive and Terrorist Content Regulation (as well as the Glawischnig-Piesczek ruling) to design and deploy automated moderation solutions, and by the Terrorist Content Regulation and Digital Services Act to use their contractual terms and conditions to forbid undesired behaviour. These regulatory devices mean that there will be little distinction between content moderation for law enforcement purposes and commercial purposes. Platforms typically remove content under their contractual terms and conditions where possible, in order to apply consistent standards worldwide, even where it would anyway have to be removed based on applicable national law. [124] Legally-mandated moderation, voluntary moderation and content curation more generally are all based on the same technical tools and classifications. [125] In practice, this means that any automated tools developed for law enforcement will likely also be deployed more widely in platforms’ voluntary and commercially-motivated content governance. [126] The increasing use of automated content moderation tools subjects all online communication to the distorting influence of platforms’ commercial goals. [127] This is likely to exacerbate the issues of overinclusivity and discrimination discussed in section 3 Tom Nichini TN @lydia cross-reference .

42

As with Schultz’s analysis of sexual harassment law, we should not only observe that content regulation is over- and underinclusive, but ask who benefits from this state of affairs. Schultz argues that bright-line rules aiming to eliminate any kind of sexual conduct resonate with corporate interests and managerialist ideologies, which aim to make workplaces maximally efficient and rational. [128] Feminists arguing for a ban on sexual harassment found it politically expedient to put forward arguments that aligned with these perspectives, arguing that harassment made female employees less productive. [129]

43

Similarly, we should question whose interests are served by the current approach to platform regulation. It is first relevant to note big tech companies’ gargantuan lobbying expenditures in the EU, which outstrip all other sectors. [130] They also influence broader academic and policy debates by funding think tanks, research centres etc. [131] Leading platforms have been willing to accept greater regulation, as long as it strengthens dominant market players and does not demand fundamental changes to their business models. [132] These lobbying and advocacy efforts are not only about getting the regulatory results that they want, but shifting regulators’ attention to the topics that are least threatening by amplifying “the criticism they can structurally live with”. [133] In this context, we should be attentive to potential ways that the orientation and priorities of European regulation might align with platforms’ interests, even if individual regulatory requirements are unwelcome and burdensome.

44

Just as the focus on individual sexual misconduct in sex discrimination law excuses businesses from considering organisational context and structural inequality, European regulation arguably gives platforms obligations that are easy for them to “live with” instead of demanding structural changes that might discourage harmful speech and create more equal and inclusive online environments. European regulation has been criticised for focusing on the content of individual posts, rather than contextual factors like platform design. [134] However, this orientation serves platforms’ interests insofar as it aligns with their current moderation practices [135], and with their commercial priorities. Irrespective of regulatory considerations, platforms have incentives to find and remove the most obviously offensive or illegal content, which is likely to repel users and advertisers. [136] They have much less incentive to redesign recommendation algorithms and platform architectures that contribute to social harms, given that these architectures in their current form are optimised for profit. In focusing on moderation at the content level rather than broader contextual, structural and design considerations, EU regulation effectively aligns with platform priorities more than the public interest.

45

It also reflects the influence of other powerful stakeholders. The new forms of private ordering that the EU has promoted in areas like terrorist content and disinformation involve close cooperation between platforms and national authorities. This not only enables those authorities to censor content online while circumventing formal legal channels and the checks and balances they entail [137], but also facilitates security agencies’ collection of data on platform users and their activities. [138] EU regulation has also been particularly heavily influenced by lobbying from the copyright industries [139]—so much so that platforms are now, rather counterintuitively, subject to stricter intermediary liability for copyright infringement than for any other type of content, including terrorist content or child sexual abuse material. [140] Copyright owners are primarily interested in restricting the availability of specific content in which they have an economic interest, not in broader considerations about how online environments are constructed. This natural tendency towards a content-level orientation in one of the EU’s highest-priority policy areas may have influenced its approach in other areas of social media regulation: an example is the notice and takedown system, which was originally developed in the US Digital Millennium Copyright Act but now applies to all illegal content in the EU. [141] Thus, the over- and underinclusivity of the EU’s platform regulation seems to reflect the interests of a variety of state and corporate actors in prioritising content-level regulation and surveillance over designing safe and egalitarian online spaces.