🍻 Selber Bier Brauen

In 2011, Marc Andreessen, a key figure in California’s venture capital scene, coined the phrase: “Software is eating the world.” The phrase describes the spread of software into everyday life and the displacement of physical business models. This process continues in an unexpectedly literal sense: AI companies purchase used books, scan them, and dispose of them to gather input for their models. Anthropic disclosed this practice earlier this year during a court proceeding. As the Washington Post reports: “The document describes how the scanning company’s ‘hydraulic powered cutting machine’ would ‘neatly cut’ books, whose pages would later be ‘scanned on high speed, high quality, production level scanners.’ Finally, it notes, the scanning company will ‘schedule with the recycling company to pick up the completed books.’” In Europe, reports are mounting regarding the purchasing practices of the Canadian company Zoom Books: The online bookstore is acquiring hundreds of thousands of nonfiction titles for which there is no longer a large market, but which represent previously unused material for training models.

The reason for this seemingly cumbersome method is the expectation that it will fall under the fair use provision of U.S. copyright law. It is a new stage of the AI copyright war: There is no flexible instrument such as the fair use doctrine in EU copyright law, but the opt-out structure of the exception and limitation for data mining might make used books fair game. Does one emphasize the creative or the destructive aspect of what Schumpeter called “creative destruction”? For some, book-devouring AI training signifies a capitalist fever dream: hungry machines exploiting the work of hundreds of thousands of creators and the cultural heritage of their readers, their billionaire overlords working towards a cyberpunk future known from books like Neuromancer or video games like Cyberpunk 2077. But concerns about the use or abuse of old books are also an indication of the problems facing European culture: The problem is not that Americans buy unwanted books, but that books are more and more unwanted in Europe.

Used Books as Fair Game?

Ownership of a physical copy of a book does not confer the right to reproduce it, neither in the U.S. nor in the EU. But the respective copyright laws use different instruments to balance the interests of rightholders and the public. In the US, this balance hinges on the fair use doctrine  (Section 107 Copyright Act). Whether use without permission is fair depends on four factors, with a strong focus on the questions of how similar the use is to the work being used and whether they compete economically. In the EU, the right to use copyrighted material without the rights holder’s consent depends on whether such use falls under one of the exceptions and limitations set forth in Article 5 Infosoc Directive. It gives preference to certain forms of use but does not allow for an exception on a case-by-case assessment beyond the written catalogue: a gain in legal certainty that comes at the cost of less fairness and flexibility in individual cases.

The debate over the legal limits of book digitization predates AI – it began with Google Books. Google scanned more than 40 million titles without the consent of the authors, although the company later implemented an opt-out clause. The company scanned works in the public domain – whose copyright protection had expired – as well as works that were still fully protected by copyright. Google shows previews of the books online, the corpus also functions as a basis for their Ngram Viewer. Google’s action was adjudicated in Authors Guild, Inc. v. Google, Inc. and deemed to be fair use: “Google’s unauthorized digitizing of copyright-protected works, creation of a search functionality, and display of snippets from those works are non-infringing fair uses. The purpose of the copying is highly transformative, the public display of text is limited, and the revelations do not provide a significant market substitute for the protected aspects of the originals. Google’s commercial nature and profit motivation do not justify denial of fair use.” While things had quieted down around Google Books, AI companies’ hunger for data has given the topic new significance. AI companies are searching for unused corpora and are well known for having resorted to reckless methods in their search for materials: for example, by using corpora from platforms with questionable copyright compliance, such as Libgen or Anna’s Archive. Recent cases have been decided in favor of Meta and Anthropic. However, these decisions do not amount to a free pass, regardless of how the companies obtained the books. And, more importantly in practical terms: digital pastures have been grazed – physical books are their last frontier.

This raises the question of how the same practice fares under European law: Is the use of copyrighted works, whether digital or physical, illegal in Europe? The digitization of the book (as well as later copies in the learning process) constitutes a reproduction – one of copyright law’s exploitation rights. It is only permitted if it falls under one of the exceptions and limitations of the Infosoc Directive. But there is no such exception or limitation. Alas, Art. 4 of the DSM Directive introduced an exception and limitation for text and data mining. It is now generally accepted that AI training falls under the definition of data mining. Its central feature is that it guarantees creators an opt-out. If they object to the use, the exception and limitation is no help. Anyone who opens a newly printed book will find an opt-out at the beginning. In the case of old books that the AI companies digitized themselves, neither the technological prerequisites for data mining nor the legal option to opt out existed. Therefore, an opt-out is not declared. This applies not only to truly old books, but also to those that are only a few years old but were printed before the DSM Directive introduced the exception and limitation for data mining. Are used books fair game? According to Recital 18, it is possible to declare the opt-out not in the copy but by other means. This gives publishers the ability to prevent even books that have already been published before the DSM Directive from being used for data mining. But some publishers have gone out of business and some authors have passed away. Not every author has a legal department to look after them. Without an opt-out, it all depends on the unclear meaning of the work being “lawfully accessible” (Art. 4 (1) DSM Directive). A wide interpretation is possible on the level of the DSM Directive and its national implementation in Germany. But its compatibility with European and national fundamental rights does not seem self-evident: after all, this involves a significant encroachment on authors’ rights.

The Whimpering of Hollow Men

In America, old books are midwifing the machine god. What about Europe? Lamenting the use of old books as abuse feels like compensatory behavior: Europeans begrudge American machines the books in which they themselves have long since lost interest. We are witnessing an unmitigated crisis of literary culture in publishing, education and politics.

A few examples illustrate the changes that have taken place since the middle of the last century: after Thomas Mann wrote Doctor Faustus while in exile in California, he went on to write a commentary on its creation – a miniature of the literary culture of modernity in which art was high art: difficult, demanding, grand. Today, love stories between werewolves and vampires popularized on TikTok, or cross-selling fantasies like Save Me/Maxton Hall, may generate revenue for publishers and some self-publishers in the young adult category. But only from a philistine’s perspective does reading have value as such. The continued existence of books as a medium masks the cultural rupture. A rupture aided and abetted by the state. Two examples from Germany: students’ reading proficiency continues to decline; according to the IGLU study, one-quarter of elementary school students do not meet the minimum standard. At the same time, classics are being distorted in Berlin schools by being read in so-called “simple language.” This decline in people’s intellectual capacity is reflected by their representatives: during a visit to Colombey-les-Deux-Églises, Konrad Adenauer asked President de Gaulle if he could spend an hour alone in his library – to get to know the French politician. Today, former chancellors like Angela Merkel or Olaf Scholz are considered well-educated if, while traveling abroad, they tell journalists that they are reading a book on the history of the country they are visiting. Meanwhile, Thuringia’s Minister President, Mario Voigt, is reported to have his speeches partly generated by AI – even on an issue like the Holocaust. Why bother? He leaves the reading of books to American machines.

The post AI Is Eating the Book World appeared first on Verfassungsblog.

As AI-generated media flood our information ecosystems, detecting synthetic content has become an urgent regulatory challenge – in fact, not one challenge but many, as synthetic media breeds problems across a range of digital contexts, including deepfakes and disinformation, scamming, and content moderation. The EU’s new “Code of Practice on Transparency of AI-Generated Content” – the first concrete articulation of Article 50(2) AI Act, the EU’s approach to AI-content detection – gives sensible answers to several open questions and will advance the global regulatory debate. Unfortunately, however, it cannot escape the structural shortcoming at the core of Article 50(2): the premise of “general-purpose AI detectability,” which treats detection as a “fixable” technical problem rather than a context-dependent, political one demanding iterative, multi-actor coordination. That diagnosis points towards a different response: rethinking synthetic-media regulation, and kindred problems of internet governance, around institutionalized, regulator-led forums that oblige stakeholders to keep communicating and collaborating.

Synthetic Content Transparency: Article 50 AI Act and the EU’s new Code of Practice

Establishing AI detectability is no simple problem: it has to deal with a quickly evolving technological landscape and a range of diverse actors – AI developers, platforms, researchers and more – whose interests only partly overlap. Fortunately, regulators have nonetheless begun to act, adopting legislation to help increase synthetic media detection capabilities. One of these regulators, alongside China and India, is the EU, which, in Article 50(2) AI Act, obliges “[p]roviders of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, [to] ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated”. Furthermore, “[p]roviders shall ensure their technical solutions are effective, interoperable, robust and reliable as far as this is technically feasible”.

With the adoption of the “Code of Practice on Transparency of AI-Generated Content” on 10 June 2026 – a guidebook developed by appointed, yet independent academic experts together with over 100 stakeholders from industry, academia and civil society – these obligations have now found a concrete articulation. The Code of Practice lays down a framework of commitments and measures – some of which obligatory, some optional – that can prove adherence to Article 50(2). If the Commission approves the Code, which seems likely, adherence to the Code will create a presumption of compliance with Article 50(2). Although providers are also free to adopt their own, alternative compliance strategies, the legal certainty provided by this presumption will in all probability confer high practical relevance on the Code.

As the remainder of this contribution argues, the Code, for all it gets right, exhibits a number of questionable blindspots and, perhaps inevitably, fails to overcome the structural shortcoming at the core of Article 50(2).

What the Code gets right

One aspect that will be central to the success of any AI detectability strategy is the issue of technical implementation. Synthetic media can only be detected if appropriate technical tools are available to successfully identify AI generations. The issue is a complex one: it is unlikely that concerned actors will ever possess technologies that can ensure a robust marking of all covered forms of synthetic media; as of today, they don’t. The Code obliges AI providers to implement at least two layers of machine-readable marking, focusing on the currently dominant approaches of metadata addition and imperceptible watermarking. However, it also calls on providers to continually test and overhaul their tools to at least attempt to maintain effectiveness and robustness vis-a-vis evolving adversarial challenges – a generally convincing approach.

The Code also takes several welcome steps to advance the coordination between AI developers and other actors involved in content propagation and consumption, which is pivotal to the (inherently collaborative) project of synthetic media detectability. These include the duty for AI providers to prohibit users, through their acceptable use policies, from intentionally removing or tampering with markings and encouraging them to “collaborate with relevant actors within the ecosystem to make their detection solution directly available within distribution and communication platforms”. While this does not reach the level of integration implemented by China’s regulation on AI content transparency, which extends marking and labelling duties to providers of generative AI systems as well as platform intermediaries and end users, it still marks a significant, if tentative, step toward the cross-actor coordination on which synthetic media detectability depends.

The blindspots

Unfortunately, the Code also suffers from a number of blindspots. First of all, it insufficiently considers that any successful AI detection strategy needs to have special consideration for the problem of malicious actors. Whereas AI slop and other “innocent” synthetic media propagation have risks, the dangers posed by malicious actors willing and able to strategically bypass detection mechanisms (e.g. for disinformation) are not only more grave, but also harder to mitigate. Synthetic media identification, therefore, is structured by what has adequately been called a “detection dilemma”: the more broadly accessible a detection technology becomes, the more easily it can be circumvented. This dynamic is compounded by an asymmetry of cost and effort: producing convincing synthetic media has already become incredibly cheap and fast, while marking and detecting it remains technically demanding and expensive. While the Code is not blind to this dynamic and the related insight of an “AI detection arms race”, its default is broad public availability: detection solutions are to be made available to the general public, free of charge, with encouragement toward open specifications. Staggered access, i.e. restricting detection to (verified) high-stakes users while accounting for security and equity concerns, is reserved for the case of free-form text watermarking. The result is a posture of “default public access” for the high-fidelity modalities (photo, video, audio) where determined bad actors stand to gain the most.

What is more, the Code prematurely discounts the value of passive, forensic AI detection tools, by treating them as a purely optional measure rather than a potential compliance necessity.

The Code is right in highlighting the current unreliability of forensic AI detection tools and their inferiority to active marking technologies. However, especially against malicious actors, which can strip synthetic media of embedded detection signals or make use of systems that do not apply them in the first place, forensic detection tools can be the only remedy. Forensic detection arguably fits poorly within the AI Act’s provider-centric design: developing synthetic content identifiers, in particular model-agnostic ones, is a duty one could just as readily locate with content propagators, as is the case under India’s new intermediary-centric regulation on AI detection. However, as actors with highly privileged insight into the characteristics and telltales of AI generations, AI systems providers, especially large ones, are well-equipped to share in the responsibility to build effective forensic detection tools.

A deeper misconstruction – and a way forward

These shortcomings ultimately reflect a deeper misconstruction at the core of the AI Act’s approach to synthetic media detectability. Article 50(2) prescribes a horizontal, uniform, one-sided, risk-agnostic, technical “fix” to what in reality are diverse issues of digital deception, the responses to which are necessarily context-dependent and “political” and require collaboration and coordination between a number of different actors, including AI systems providers, intermediaries, public authorities, researchers and NGOs. The Code even attempts to inspire such networking – encouraging AI providers to “foster collaboration” with deployers and users of their systems – but of course is formally reined in by the silo of “AI regulation”. Whether these synapses will be established by other regulations, e.g. the DSA’s obligations on platform and search engine providers, is an open question, but seems dubious. Unfortunately, the structure of Article 50(2) and its too broad, and therefore too narrow, perspective of ‘general-purpose AI detectability’ seem to have prevented the Code of Practice from engaging with the work that actually matters: sensibly delineating which forms of AI-generated content are genuinely deceptive, and devising the nuanced, socio-technical responses capable of countering them.

These insights might help inform future legislation. They also point to a potentially more effective and sustainable method for addressing multi-actor digital governance issues, namely one that recognizes that many such issues, including synthetic media harms, cannot be discharged once and for all by a discrete techno-regulatory intervention, but rather require continuous and evolving multi-party collaboration. Such a framework could be set up through an institutionalized, regulator-led forum that obliges the relevant actors to communicate and collaborate, and in which decisions regarding e.g. the development of and access to detection technologies are regularly revisited in light of evolving real-world problems and priorities. The Code makes a first, rudimentary step in that direction, by announcing a “taskforce … to facilitate … collaboration between Signatories, other actors in the value chain and relevant stakeholders through regular meetings”, which Signatories are “encouraged” to participate in. Yet, if the preceding analysis is correct, such a forum belongs at the core of synthetic media regulation, not in an optional taskforce.

The post The Problems with “General Purpose AI Detectability” appeared first on Verfassungsblog.

The recent judgment Miladze v Georgia of the European Court of Human Rights concerning a TikTok video published by a Georgian self-defined civil activist adds another layer to the Court’s increasingly messy Article 10 case law. The applicant repeatedly insulted public officials in crude and sexually explicit terms while broadcasting to a large online audience. Domestic courts imposed only a modest administrative fine, later reduced on appeal. The ECtHR did not find a violation of the applicant’s right to freedom of expression.

The judgment therefore raises several questions about the limits of legitimate political speech online. This post first outlines the facts and reasoning of the case before turning to a critical analysis.

The Facts

The applicant, an active social media figure working as a food deliverer in Tbilisi, published a TikTok video at the end of a long shift criticising the city mayor’s office over urban transport policies and the alleged misuse of bus lanes by public officials driving government cars. Throughout the video, he used repeated vulgar and sexually explicit insults directed at the mayor, municipal staff and law enforcement officers, including phrases equivalent to “f* you” and “go f* your mothers”.

Georgian authorities charged him under administrative law provisions concerning disorderly conduct and insults against law enforcement officers in public. According to the government, the offence had been committed in a “public place”, namely TikTok, and the obscene language used did not merit protection under freedom of expression guarantees.

The applicant argued that his statements formed part of political criticism concerning matters of public interest and were not directed at identifiable individuals. The domestic courts rejected these arguments. The first-instance court accepted that TikTok constituted a public space for the purposes of administrative liability. While acknowledging that public officials must tolerate sharp and emotionally charged criticism to a greater extent, it nevertheless concluded that direct insults fell outside protected expression. It further stressed the importance of preventing the normalisation of sexually explicit insults on social media platforms popular among minors. On appeal, the applicant’s fine was reduced, but the conviction itself was upheld.

The Judgment

Before the Strasbourg Court, the applicant argued that the interference with his freedom of expression was not prescribed by law because Georgian legislation regulating conduct in public spaces did not explicitly extend to cyberspace. He further contended that the interference lacked a legitimate aim and was unnecessary in a democratic society, particularly since the speech did not amount to hate speech or incitement to violence.

The Court quickly disposed of the first two elements of the Article 10 test. It considered it sufficiently foreseeable that social media platforms could qualify as public spaces under domestic law. Likewise, it accepted that the interference pursued legitimate aims, namely the protection of morals, public order and the rights of others. The real issue therefore concerned whether the interference was “necessary in a democratic society”.

Here, the Court repeated familiar formulas from its Article 10 jurisprudence, such as that protection under art 10 ECHR extends to expressions that “shock, offend or disturb” and that vulgarity does not automatically deprive speech of Convention protection, especially where such language serves a stylistic or expressive purpose in debates of public interest. At the same time, it also reminded that national authorities enjoy a certain margin of appreciation when balancing competing interests (§ 67-69).

The Court nevertheless sided with the Georgian authorities. Its analysis focused heavily on the nature of the applicant’s speech, the personalisation of the statements, and the characteristics of TikTok as a medium.

Provocation vs Denigration

According to the Court, substantial portions of the video consisted not of political argument but of sustained verbal aggression devoid of informational value. While the applicant’s criticism originated in a matter of public interest, namely urban transport policy and alleged official misconduct, the Court considered the form of expression decisive. The language used was characterised as personally abusive and “wanton denigration” rather than politically provocative (§ 71-72).

This sits uneasily alongside earlier Strasbourg case law strongly protecting offensive political expression. In Peradze and Others, the Court explicitly recalled that Article 10 protection extends even to “clearly vulgar and/or offensive language and/or conduct, including expressions with sexual references”, provided that such language contributes to public debate on matters of public interest (§ 45).

In my view, the present judgment fails to explain convincingly why that principle suddenly loses force here. The applicant’s statements, however crude, emerged directly from criticism of local government policy and alleged abuses of authority. This was a citizen’s angry denunciation of public officials and urban governance, not a mere insult detached from political debate. Yet the Court artificially separates the political content from the emotional and vulgar form through which that criticism was expressed. Nor is the Court’s distinction between protected provocation and unprotected “wanton denigration” sufficiently clear (§ 68). The circumstances here clearly differ from those in Rujak v Croatia, where the standard of “wanton denigration” was applied to statements made during a quarrel with public officials with the sole intent of offending them and without any connection to a matter of public concern.  The present judgment understates the public relevance of the applicant’s speech by dismissing any of its expressive or informational value, while offering little reasoning for that conclusion beyond moral disapproval of the language itself.

Individual Attack vs Critique of Authority

The Court takes care to distinguish the case from Peradze and Others v Georgia, where the applicant chanted vulgar statements at a public protest against a construction project without targeting specific individuals. In the present case, by contrast, the insults targeted the mayor and law enforcement personnel specifically (§ 72).

The Court arguably overstates the personalised nature of the attack. Although the mayor was identifiable, the impugned comments were not directed exclusively at him but also referred more broadly to municipal staff and law enforcement officers. Moreover, direct offences against an identifiable public official have been considered worthy of protection. It is unclear why the Court decided to depart from its frequent admission that “political invective often spills over into the personal sphere; such are the hazards of politics and the free debate of ideas, which are the guarantees of a democratic society” (§ 34).  The judgment therefore expands the notion of impermissible personal abuse in ways that may prove difficult to confine in future cases involving online criticism of public authorities.

The Court’s approach is also difficult to reconcile with its longstanding insistence that public officials must display a higher degree of tolerance toward criticism, even when it comes to direct insults. This was, for instance, the case in Gaspari v Armenia, where a protester called a chief police officer “scum” and in Eon v France, where a dissenting citizen referred to the President as a “prick” on a publicly displayed placard. Similarly, in the seminal case Oberschlick v Austria, a journalist’s utterance of the word “idiot” when publicly describing a politician was deemed protected. In all these and other cases, the Court has emphasised the narrow limits of acceptable restrictions on political expression directed at public officials (including municipal authorities).

In the case at hand, the applicant’s remarks were more vulgar than the expressions at issue in those cases. Yet, the Court has repeatedly asserted that vulgarity alone cannot be decisive in the assessment, given that it could serve stylistic purposes and that style enjoys protection just like the content of expression (i.e. Uj v Hungary, § 20). In Bouton v France, for instance, the Court protected a highly provocative political protest involving nudity and sexually explicit language (“f*ck the Church”). However, in the present case, the Court rejected analogies to cases involving satire, artistic provocation or hyperbolic political speech, emphasising the absence of any stylistic value and the violent and degrading implications of the insults within the Georgian linguistic context (§ 74). In my view, the application of this principle here leads to a rather artificial dismissal of outrage, provocation and emotional immediacy as the language of political contestation, particularly those, like the applicant, challenging abuse of public power from positions of social disadvantage and frustration.

Social Media and “Obscenity”

Another significant factor weighed in by the Court is the impact of viral amplification caused by TikTok’s algorithmic design. The Court also repeatedly referred to TikTok’s broad accessibility and popularity among minors. The video reportedly accumulated around 100,000 views within a short period. Although the applicant included an initial warning concerning vulgar language, the Court considered this insufficient to prevent exposure to underage audiences. In this context, it stressed the importance of preventing the normalisation of sexually explicit insults in online discourse (§ 76).

The Court’s approach here resonates with its case law for a stricter regulatory approach to online defamatory content (§ 75). In the case at hand, however, this reasoning appears to overstate the level of obscenity in the applicant’s statements, when compared to the type of harm in respect of which the Court has traditionally afforded States a wide margin of appreciation in protecting minors, such as explicit sexual advice and imagery. By considering the national authorities’ treatment of the applicant’s language as a matter of public morality requiring legal intervention, the Court edges dangerously close to conflating offensive, yet (as shown above) protected, political speech with genuine online harm. Once TikTok and equivalent platforms are acknowledged as public spaces of political debate, it becomes difficult to justify expectations of unusually sanitised political discourse within them.

Proportionate Sanctions

Finally, the Court considered the reduced administrative fine as proportionate and limited in its chilling effect, also in the absence of additional restrictions (§ 77). However, the Court forgets how, as per its case law, even the minimum statutory sanction or a warning could have this effect.

Blurred Standards

This decision contributes to the broader instability in the Court’s contemporary Article 10 jurisprudence. Strasbourg continues to proclaim robust protection for offensive political speech while simultaneously narrowing that protection in practice through elastic and ambiguous analyses. As Natali Alkiviadou has recently observed in her work on hate speech and the ECtHR, the Court’s distinctions between protected provocation and punishable abuse often remain inconsistent, ungrounded in empirical evidence and difficult to predict. Moreover, the ruling advances the blurring of standards of responsible communication, once primarily associated with professional media actors, to other speakers. Besides the tone and content, the applicant was effectively expected to account not only for his own expression but also for algorithmic amplification, audience vulnerability and the broader communicative environment of TikTok itself.

While I do not wish to overstate the relevance of this single judgment (also given that it originates from a dispute in a non-EU country), I cannot avoid expressing concern about the implications of this kind of legal reasoning beyond Article 10 doctrine. As European regulatory frameworks such as the EU’s Digital Services Act place growing emphasis on the prevention of online harms in platform governance, authoritative human rights institutions like the ECtHR may, by conflating vulgar political expression with socially harmful conduct, (indirectly) legitimise unnecessarily restrictive moderation practices. The resulting chilling effect would likely fall most heavily on informal and emotionally charged forms of civil activism.

The post The Politics of Provocation appeared first on Verfassungsblog.