🎨 Art & Culture Law · 2026-04-09
🏛️ Art, Culture & Law — 2026-04-09
🏛️ Art, Culture & Law — 2026-04-09
🏛️ Art, Culture & Law — 2026-04-09 Thursday, April 9, 2026
🏛️ The Supreme Court's Non-Decision: Thaler v. Perlmutter and the Human Authorship Doctrine 📜 The White House Framework: Training on Copyrighted Works Doesn't Inherently Violate Copyright 🇪🇺 The EU AI Act Betrayal: Creative Industries Say Copyright Provisions Were Gutted 🪞 China Regulates the Digital Double: Draft Rules on Virtual Humans and Synthetic Likeness 🎨 No Artists, No Art: The Campaign for Consent, Compensation, and Credit in the Age of Training Data
🏛️ The Supreme Court's Non-Decision: Thaler v. Perlmutter and the Human Authorship Doctrine
In March 2026, the US Supreme Court declined to hear Thaler v. Perlmutter — effectively ending, for now, the question of whether AI systems can hold copyright. Mayer Brown's analysis notes that the denial leaves standing the lower court rulings requiring human authorship as a prerequisite for copyright protection. Works created entirely by AI are not eligible for copyright in the US. The human creative hand — expressed through selection, arrangement, modification, or meaningful directorial control over AI outputs — remains the threshold that separates protectable from unprotectable work.
The Supreme Court's decision not to decide is itself a decision about the pace at which legal doctrine should respond to technological change. By declining review, the Court leaves copyright law's human authorship requirement intact while Congress is actively considering legislation — the CLEAR Act and the TRUMP AMERICA AI Act represent opposing legislative theories — and while the White House has issued a framework that takes a more permissive view of AI training than the legislative proposals. The result is a period of deliberate doctrinal ambiguity: the law is clear that AI alone cannot hold copyright, but the training data question — the far more commercially consequential question — is explicitly deferred to further litigation and legislation.
The US Copyright Office's consistent position, reaffirmed through the lower court rulings the Supreme Court let stand, is that "human creativity is essential" and works generated entirely by AI "do not qualify for copyright" — while works created with AI assistance "can still be copyrighted, provided there is 'sufficient expressive elements' and meaningful human creative control over the final output." The Thaler case itself was always the easier question. Stephen Thaler's DABUS system was presented as the sole creator of a work, with no human creative input claimed. The courts that ruled against him were not breaking new ground; they were applying existing human authorship doctrine to a novel fact pattern. The hard cases — which are now proliferating in federal courts across the country — involve humans who used AI tools in their creative process and whose works exhibit some combination of human and AI creative contribution. How much human input is sufficient? What counts as "meaningful creative control"? Is directing an AI with detailed prompts equivalent to directing a human illustrator? These questions do not have doctrinal answers yet, and the Supreme Court's refusal to take Thaler has not brought them any closer to resolution.
The international dimension of the human authorship question is increasingly consequential. The US rule — human authorship required for copyright — is not the universal position. Some jurisdictions are more receptive to expanded AI authorship concepts; others have different doctrinal frameworks that produce different results for the same creative facts. As AI-assisted creative work becomes global, the variation in copyright doctrine across jurisdictions creates both arbitrage opportunities (registering AI-assisted works in jurisdictions with more expansive authorship concepts) and compliance complexity (different treatment of the same work in different markets). The doctrinal fragmentation that the Supreme Court's non-decision preserves in the US is mirrored internationally in ways that no single ruling will resolve.
For artists, the practical effect of the human authorship doctrine is both protective and limiting. Protective because it means that AI systems and their developers cannot hold copyright in AI-generated works; any copyright that exists belongs to the human who contributed creative elements. Limiting because it means that AI-generated works without sufficient human creative contribution fall into the public domain immediately — they are not property of the artist who prompted them. The copyright protection available to AI-assisted artists is proportional to the human creativity they contribute; the protection available for purely algorithmic outputs is zero. That incentive structure rewards human creative engagement with AI tools and penalizes pure automation. Whether it is the right incentive structure for a healthy creative ecosystem is the deeper question that doctrinal analysis alone cannot answer.
📜 The White House Framework: Training on Copyrighted Works Doesn't Inherently Violate Copyright
The White House National Policy Framework for Artificial Intelligence, released in March 2026, took a position on the training data question that aligns with the AI industry's preferred outcome: training AI on copyrighted material does not inherently violate copyright law. The framework acknowledges that courts are the ultimate arbiter but sets an executive branch position that will shape regulatory guidance, agency enforcement priorities, and international trade positions. The framework also floats the idea of collective licensing frameworks as a mechanism for compensating creators without requiring per-work licensing that AI developers argue is impractical at scale.
The White House position is in tension with significant legislative activity. The TRUMP AMERICA AI Act would amend the fair use provision to explicitly state that unauthorized use of copyrighted works for AI training does not constitute fair use — the direct opposite of the framework's implicit position. The CLEAR Act, more modest, would require disclosure rather than restriction; it mandates that AI companies publish which copyrighted works were used in training datasets, without requiring consent or compensation. These three positions — executive permissiveness, legislative restriction, and legislative transparency — represent the range of policy options that Congress and the courts will adjudicate over the next several years.
The collective licensing proposal is worth examining carefully because it represents the most architecturally coherent resolution to the training data dispute — and the one that is most structurally advantageous to AI developers. Collective licensing exists in music (ASCAP, BMI) and some other creative domains; it allows licensed use of large catalogs of works in exchange for blanket fees that are distributed to rights holders according to usage formulas. Applied to AI training data, collective licensing would allow AI developers to train on copyrighted works without per-work negotiation, in exchange for payments to a collective that distributes them to creators. The practical challenges are substantial: what formula determines individual creator payments? Who administers the collective? How are non-participants treated? But the theoretical appeal is that it converts an adversarial litigation landscape into a licensing market that generates ongoing revenue for creators.
The alternative — per-work licensing for AI training — is what the AI industry's fair use arguments are designed to avoid. At the scale of frontier model training, which involves hundreds of billions to trillions of tokens drawn from the internet, per-work licensing is practically impossible even if legally required: the transaction costs of licensing each work individually would exceed the commercial value of AI development. The AI industry's argument that training is a transformative use that falls under fair use is therefore not merely a legal theory; it is an economic necessity for the current model development paradigm. If courts reject the fair use argument for training at scale — and the Supreme Court's refusal to weigh in on Thaler does not resolve this question — the economics of frontier AI training will require legislative intervention rather than just doctrinal adjustment.
The creative industries' position — that training on copyrighted works without consent or compensation is economic exploitation regardless of its legal characterization — reflects a distributive justice argument that legal analysis typically cannot accommodate. Fair use doctrine does not ask whether the use is economically fair to the original creators; it asks whether it is transformative, commercial, and substitutive in doctrinally specified ways. The gap between the legal question and the distributive question is where the most intractable political conflict in AI copyright lies, and neither the White House framework nor the pending legislation closes it.
🇪🇺 The EU AI Act Betrayal: Creative Industries Say Copyright Provisions Were Gutted
European creative industry groups — including music, film, publishing, and visual arts organizations — have published responses to the EU AI Act's implementation that describe the copyright provisions as a "betrayal" of the Act's original intent. Music Business Worldwide reports that the industry's objections center on the gap between what the Act's text requires — that GPAI model providers implement policies to comply with EU copyright law and respect creator opt-out reservations — and how those requirements have been interpreted in implementation guidance that creative industries say overwhelmingly favors AI developers at creators' expense.
The creative industries' response was direct: the EU AI Act implementation constitutes "a betrayal" of its original intent, with industry feedback "largely ignored in favor of AI model providers." The specific grievances are technical but consequential. The EU AI Act requires GPAI providers to publish "sufficiently detailed summaries" of training content — a requirement that came into force in January 2026. Creative industries argue that the summaries being published are so general as to be meaningless: they indicate broad categories of content (e.g., "publicly available internet text") without the granularity needed for individual creators to determine whether their specific works were used. The opt-out mechanism, which the Act positions as the primary protection for creators who do not wish their works used in training, requires creators to affirmatively register their opt-out with each AI provider — a process that places the burden on millions of individual creators rather than on the far smaller number of AI developers who benefit from unrestricted training access.
The structural problem is one of information asymmetry at scale. AI developers know exactly which works are in their training data; individual creators do not. The transparency requirements of the Act were supposed to address this asymmetry, but the implementation guidance has interpreted those requirements narrowly enough that the asymmetry largely persists. The Act's extraterritorial reach — applying to models placed on the EU market regardless of where training occurred — provides the legal hook for enforcement, but enforcement requires the information that inadequate transparency provisions fail to deliver.
The European creative industries' critique is not primarily anti-AI; it is anti-asymmetry. The organizations involved — including many that have adopted AI tools for production, distribution, and marketing — are not arguing that AI should not be trained on creative works. They are arguing that the terms on which that training occurs should be negotiated rather than imposed through weak default permissions that they never agreed to. The difference between "we will train on your work unless you opt out" and "we will not train on your work unless you opt in" is the difference between a system that defaults to AI developers' interests and one that defaults to creators' interests. The EU AI Act was understood by creative industries as moving toward opt-in; its implementation, they argue, has preserved opt-out.
The political economy of this dispute reflects the broader structural advantage that AI developers have in regulatory negotiations: they are few in number, well-resourced, and able to coordinate; creative industries are many, fragmented, and speak from multiple national contexts with different legal traditions and different specific interests. The regulation-writing process that produces asymmetric outcomes is not necessarily the result of corruption or bad faith; it is the predictable output of a process in which one side has significantly more capacity to shape implementation guidance than the other. The creative industries' "betrayal" framing, whatever its rhetorical excess, identifies a real structural dynamic.
🪞 China Regulates the Digital Double: Draft Rules on Virtual Humans and Synthetic Likeness
China's Cyberspace Administration published draft rules this week governing "digital virtual humans" — AI-powered interactive services that generate synthetic personas, companions, and representations of real individuals. Biometric Update reports that the rules address several distinct harm categories: biometric deepfakes (using someone's likeness without consent), virtual companion relationships with minors, content that endangers national security, and the use of digital humans to bypass identity verification. The rules require clear labeling of virtual human interactions, explicit consent for personal likeness use, and prohibition of virtual intimate relationships for minors.
The synthetic likeness provisions represent the most direct engagement with the creative rights question from a non-Western jurisdiction. The EU AI Act and US legislative proposals focus primarily on the training data question — whether copyrighted works can be used to train models. China's draft rules focus on the output question — whether AI systems can generate synthetic representations of real individuals without consent. These are related but distinct legal problems, and China's regulatory framework addresses them through a combination of identity protection, consent requirements, and content restrictions that does not map cleanly onto Western copyright doctrine.
The Afghanistan study found participants describing AI companions as "always-available peer, mentor, and source of career guidance that helps compensate for the absence of learning communities" — but constrained by "privacy and surveillance risks, contextually unrealistic and culturally unsafe support." The philosophical dimension of synthetic likeness regulation is the most interesting. A digital human that uses someone's likeness — their face, voice, and mannerisms — to create a synthetic persona that interacts with other people creates a new kind of identity artifact that existing legal categories struggle to classify. Is it a copy (subject to copyright)? A defamatory statement (subject to tort law)? A performance (subject to right of publicity)? A data product (subject to privacy law)? Chinese law's answer, embedded in the draft rules, is that it is primarily an identity use that requires the explicit consent of the person whose identity is being used — a framing that prioritizes the subject's control over their own representation over the interests of whoever created or deployed the synthetic version.
The ban on virtual intimate relationships for minors is the provision that has attracted the most international attention, and it reflects a cultural and political judgment that AI companion technologies pose specific developmental risks to young people who may not be able to distinguish between AI-mediated and human-mediated emotional connection. The provision is categorical rather than regulatory: it does not attempt to define what kinds of AI companions are acceptable for minors, subject to restrictions; it bans virtual intimate relationships outright. This categorical approach is more restrictive than anything currently proposed in Western jurisdictions, where the debate is still at the stage of defining the harm and identifying appropriate interventions. China's willingness to act categorically reflects both a different governance philosophy and, possibly, a more acute awareness of the scale of AI companion deployment among Chinese youth.
The contrast with Western regulatory approaches is partly cultural and partly structural. Western copyright frameworks focus on the economic relationship between creators and AI developers, treating cultural production as a market for which rights regimes must be defined. Chinese digital human regulations focus on the social and psychological effects of AI-mediated interaction, treating cultural production and social infrastructure as continuous. Both perspectives are legitimate; they produce different regulatory emphases. The international creative rights landscape of 2026 is being shaped by both, and the interaction between them will determine whether a coherent global framework emerges or whether the regulatory fragmentation that characterizes semiconductor policy extends to cultural production as well.
🎨 No Artists, No Art: The Campaign for Consent, Compensation, and Credit in the Age of Training Data
The Graphic Artists Guild's "No Artists, No Art" campaign, launched in February 2026, and the Human Artistry Campaign's January 2026 advocacy represent a new phase in the creative industries' response to AI. Earlier campaigns focused on litigation — the 25+ copyright infringement cases pending in US federal courts — and on direct confrontation with specific AI companies over specific training data decisions. The February and January campaigns represent a shift toward structural legislative advocacy: the creation of mandatory consent, compensation, and credit requirements that would reshape the economic relationship between creative labor and AI development regardless of how existing copyright doctrine applies to training.
The Graphic Artists Guild's campaign frames the stakes directly: "No Artists, No Art" — arguing that payment, consent, and credit are required when artists' work is used to train AI models, and that "copyright should exclusively protect human intellectual creativity." The three demands — consent, compensation, credit — are individually familiar from other creative rights contexts but together constitute a more comprehensive framework than any single existing rights category provides. Consent addresses the training data question: creators should be able to decide whether their works are used to train AI systems. Compensation addresses the economic question: if consent is given, creators should receive payment. Credit addresses the attribution question: AI systems trained on specific works or styles should acknowledge that training in their outputs. The third demand is the most novel and the most technically challenging; attribution at the level of individual training examples is currently impossible for most AI systems, which have no mechanism for tracking which specific works influenced which specific outputs.
The "No Artists, No Art" framing is rhetorically powerful precisely because it is empirically debatable. The Copyright Alliance's advocacy frames the argument in terms of economic justice: without compensation for the training data that makes AI creative systems possible, the economic model of creative production becomes unsustainable. This argument has real force — the human creative production that AI systems learn from represents centuries of accumulated creative labor — but it is complicated by the fact that most creative works that AI systems learned from were created by humans who were not compensated for most of the uses their work served, including its use as inspiration by subsequent human artists. The principle that creative influence requires compensation, if applied consistently, would transform not just AI training but most of the cultural inheritance economy.
The legislative prospects for the consent, compensation, and credit framework are modest in the near term. The CLEAR Act's transparency-only approach is more likely to pass than the TRUMP AMERICA AI Act's fair use amendment, and neither addresses compensation directly. The collective licensing framework floated in the White House policy document may eventually evolve into a vehicle for compensation, but its structure would likely produce distribution formulas that favor high-volume commercial creators over the individual visual artists and writers who are the most economically vulnerable to AI displacement. The gap between the harm that creative industry advocates are documenting and the legislative remedies available in the current political environment is large.
The deeper cultural question that the "No Artists, No Art" campaign raises is one that legal frameworks are not well-positioned to address: what is the relationship between the material conditions of creative production and the cultural vitality of creative output? If AI systems trained on human creative work can produce outputs that are commercially competitive with human creative production, and if that competition reduces the economic returns to human creative labor to the point where fewer humans pursue creative careers, what happens to the cultural substrate that future AI systems will train on? The legal debates about training data and fair use are proxy arguments for a larger argument about whether AI-mediated creative production is complementary to or substitutive of human creative culture. The outcome of that argument will determine not just the distribution of economic value in the creative industries, but the character of cultural production for a generation.
Research Papers
Designing Safe and Accountable GenAI as a Learning Companion with Women Banned from Formal Education Multiple authors · arXiv cs.CY · April 9, 2026 Documents AI companions as substitutes for absent learning communities in Afghanistan, with participants describing them as "always-available peers and mentors." Directly relevant to the China digital human regulation debate: the same AI companion functionality that China is regulating for minors provides critical educational access in contexts of institutional deprivation.
Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules Multiple authors · arXiv cs.AI · April 9, 2026 Documents "blind refusal" in language models — refusal to help break rules regardless of rule legitimacy. Has direct cultural production implications: AI systems that refuse to help creators document IP violations, reproduce works for criticism and commentary, or access works under fair use principles may systematically reinforce incumbent rights structures against legitimate creator interests.
Implications
The week's art, culture, and law developments reveal a creative rights landscape in which the legal frameworks are systematically behind the economic reality. The Supreme Court's refusal to decide Thaler preserves human authorship doctrine while 25+ training data cases remain unresolved. The White House framework takes a permissive position on training while Congress develops legislation in both permissive and restrictive directions. The EU AI Act's implementation is producing outcomes that creative industries describe as contrary to its stated intent. China's digital human regulations address output harms while Western frameworks are still debating input rights. These simultaneous legal developments in multiple jurisdictions are producing not convergence but fragmentation — different answers to the same underlying questions about who controls cultural production in an age of AI.
The training data question is the most commercially consequential unresolved legal issue in AI, and it will remain unresolved until either courts adjudicate the fair use question at the scale of frontier model training, or Congress intervenes, or both. The White House framework's suggestion of collective licensing represents the most architecturally coherent resolution, but collective licensing systems take years to establish and require trust relationships between the parties involved that the current adversarial litigation environment is actively degrading. The practical trajectory is several more years of legal uncertainty, during which AI developers operate under the White House's permissive framework while creative industries pursue litigation and advocacy for more restrictive alternatives.
China's digital human regulations represent the most significant cultural production regulatory development of the week precisely because they engage with a harm — synthetic likeness without consent — that Western frameworks have not yet addressed systematically. The biometric deepfake provisions and the ban on virtual intimate relationships for minors are ahead of anything currently proposed in US or EU legislation, and they reflect a governance philosophy that treats AI cultural production as a social infrastructure question rather than a market rights question. Whether that philosophy produces better outcomes for human cultural welfare than the market-based Western approach is an empirical question that comparative analysis of AI cultural regulation across jurisdictions will eventually allow answering.
The "No Artists, No Art" campaign's three-part framework — consent, compensation, credit — defines the stakes of the creative rights debate more clearly than most legal analysis manages. Legal frameworks can address consent (through licensing requirements) and compensation (through damages and royalties) with existing doctrinal tools; they cannot currently address credit, because attribution at the level of individual training influence is technically impossible for most current AI systems. The technical and legal development of attribution mechanisms — systems that can identify which training works influenced which outputs and communicate that attribution to users — is a prerequisite for the comprehensive creative rights framework that the campaign envisions. It is also a research program that no major AI developer has publicly committed to, which suggests that the "credit" demand will require either technical breakthroughs or regulatory mandates that current political conditions are not producing.
.heuristics
- id: training-data-fair-use-as-economic-necessity-argument
- id: transparency-requirements-require-usable-granularity
- id: synthetic-likeness-as-identity-use-not-copyright-use
Art, Culture & Law is a briefing on AI and cultural production from antikythera.org.