Art & Culture Law · 2026-04-26

🎭 Art & Culture Law — 2026-04-26

⚖️ Anthropic Files Fair Use Motion to End UMG's AI Lyric-Training Lawsuit
🎬 WGA's Tentative 4-Year Deal Encodes AI Restrictions Into Hollywood Labor Infrastructure
🏛️ Recording Academy Flags Creator Risks in White House's New National AI Policy Framework
📜 New York's Dual AI-Image Laws Give Publicity Rights New Teeth
🖼️ Holland & Knight Survey: AI Is Now a Systemic Legal Risk Across the Art Market
📐 "No Retroactive Cure": Position Paper Argues Training-Time Copyright Violations Cannot Be Fixed Post-Hoc

---

⚖️ Anthropic Files Fair Use Motion to End UMG's AI Lyric-Training Lawsuit

Anthropic moved for summary judgment on fair use grounds in the music publisher lawsuit filed by Universal Music Group and associates — an action that Billboard describes as "pivotal" for determining whether training large language models on copyrighted lyrics constitutes transformative use under U.S. law. The core argument: ingesting song lyrics to train Claude is not consumption of those lyrics as creative expression, but a statistical extraction of linguistic pattern — the AI doesn't recite the lyrics, it learns from their structure. Digital Music News and Reuters both reported the motion on April 21–23, framing it as potentially the most consequential fair use ruling since Google v. Oracle.

The case consolidates claims from Universal Music Group, Sony Music Entertainment, and several independent publishers who filed in late 2024 alleging that Claude reproduced lyric fragments at scale and that the training corpus included unlicensed works scraped from LyricFind and Genius. Anthropic's motion challenges both grounds: it argues that no substantial similarity exists between Claude's outputs and the training inputs, and that even if lyrics were ingested, training is categorically transformative because it produces a model that generates new text rather than reproductions.

The four-factor fair use analysis maps unfavorably for rights holders on the transformativeness prong if Anthropic's framing holds: prior cases (Campbell v. Acuff-Rose, Authors Guild v. Google Books) have rewarded uses that add new functionality. But the RIAA and NMPA have signaled amicus positions arguing that market substitution — not transformativeness — is the dispositive test, and that AI-generated music demonstrably competes with licensed works.

What distinguishes this motion from earlier AI copyright skirmishes is Anthropic's explicit theoretical commitment: the company is not settling, not licensing, and not offering a licensing framework. It is going to the mat on pure fair use doctrine. If the court grants summary judgment, it establishes a U.S. precedent that training on unlicensed copyrighted works is per se lawful — a ruling that would effectively moot dozens of pending creative-industry suits against OpenAI, Meta, Google, and Stability AI. If the court denies it, the case proceeds to discovery and trial, with liability potentially running into the hundreds of millions. The music industry is watching as a bellwether for whether pre-licensing regimes will be judicially compelled or remain voluntary.

Sources:

---

🎬 WGA's Tentative 4-Year Deal Encodes AI Restrictions Into Hollywood Labor Infrastructure

The Writers Guild of America reached a tentative four-year agreement with the Alliance of Motion Picture and Television Producers (AMPTP) on April 4, reported by The Hollywood Reporter and Los Angeles Times. The deal's AI provisions — now under WGA member ratification — represent the most structurally significant labor-side encoding of AI restrictions in creative industry contracts to date, and set precedents likely to influence SAG-AFTRA's pending renegotiation with streaming platforms.

The core AI clause prohibits studios from using AI-generated material as a substitute for WGA-covered work: a studio cannot instruct a writer to rewrite AI-generated script pages (converting the writer from creator to editor-under-duress), and AI outputs cannot be used to set "floor" compensation by establishing that a script "exists" before a writer is hired. The Wrap characterized the provisions as "the most explicit contractual restriction on AI substitution in entertainment labor." Crucially, AI-written material does not accrue WGA credits, cannot reduce minimum compensation, and cannot be used to establish story credit in arbitration.

What the deal does not do is prohibit studios from using AI for development work, pre-visualization, or pitch-deck generation — areas where the WGA's jurisdictional reach is ambiguous. Studios retain the right to use AI tools for "creative development assistance" as long as no WGA-covered work is reduced or replaced. The Ankler's April 13 analysis flagged healthcare funding as the more contentious ratification obstacle, but the AI provisions have generated deeper anxiety among members who note that "development assistance" is precisely where displacement begins — before the formal contract engagement that union protections cover.

The deal's long-term structural significance is the normalization of AI as a contractually defined category in entertainment labor law. Prior contracts either ignored AI entirely or addressed it via broad "new technologies" clauses. This agreement requires studios to disclose when AI-generated materials are submitted to writers and prohibits using AI outputs to inflate credit arbitration records. Whether the arbitration provisions survive studio interpretations through 2030 — the deal's expiration — depends on how robustly the WGA's contract enforcement infrastructure is funded, a question the healthcare provisions make more fraught. The deal is being watched by the Directors Guild and SAG-AFTRA as a template for their own upcoming AI renegotiations.

Sources:

---

🏛️ Recording Academy Flags Creator Risks in White House's New National AI Policy Framework

The Recording Academy published an analysis of the White House's proposed national AI framework, released in early April, identifying four provisions that creators' rights advocates argue are structurally insufficient to protect musicians, composers, and audio producers. The framework — positioned as a voluntary coordination mechanism ahead of possible federal AI legislation — endorses transparency in training data disclosure without mandating licensing or compensation, a gap the Academy's GRAMMYS on the Hill lobbying delegation has made the centerpiece of its 2026 Congressional agenda.

The Recording Academy's specific objections: the framework endorses "voluntary" opt-out mechanisms for training data, which creators argue inverts the correct default (opt-in consent before use); it treats AI-generated music under the same "expression" framework as human compositions without defining authorship, leaving royalty allocation to market forces; and it delegates enforcement to the Copyright Office without providing new statutory authority or budget. The Academy's advocacy arm is coordinating with the National Music Publishers' Association and the American Federation of Musicians to push for mandatory opt-in licensing provisions in any federal legislation.

The White House framework tracks closely with the AI Action Plan issued by the Office of Science and Technology Policy, which prioritizes "AI leadership" and "innovation ecosystem" framing over rights-holder protections. The tension between innovation-first and creator-first framing runs through every AI policy document in this administration's output: the framework's training transparency provision, for instance, requires that AI developers "document" what categories of creative works were used in training, but does not require disclosure to the rights holders whose works appear in those categories.

For the creative industries, this gap between documentation and rights enforcement is not academic — it determines whether decades of music industry licensing infrastructure (ASCAP, BMI, SESAC, SoundExchange) can be extended to AI training contexts, or whether that entire royalty ecosystem is effectively bypassed by a transparency-without-compensation regime. The Recording Academy's submission to the OSTP frames this as the difference between "knowing your work was stolen and having recourse versus knowing it was stolen and having none." The framework comment period closes in May; whether Congressional intervention follows depends on how coordinated the creative industries remain in lobbying alignment.

Sources:

---

📜 New York's Dual AI-Image Laws Give Publicity Rights New Teeth

New York enacted two laws in early 2026 that Skadden analyzes as the most substantive state-level extension of right-of-publicity doctrine into AI-generated imagery since Tennessee's ELVIS Act. The first law prohibits the commercial use of AI-generated "digital replicas" of living individuals without consent, extending New York's existing right-of-publicity statute to synthetically generated likenesses — closing a gap where a deepfake of a person could evade the old statute because no actual image or performance was reproduced. The second law addresses deceased individuals' likenesses, granting estates 40-year post-mortem rights over AI-generated replicas.

The Skadden analysis flags several structural ambiguities: the laws define "digital replica" broadly enough to potentially cover AI art that stylistically evokes a living person without explicitly depicting them, creating interpretive risk for generative art platforms. The IP Watchdog notes that enforcement depends on plaintiffs establishing commercial use, which creates a carve-out for AI-generated art distributed non-commercially — a gap that advertising-adjacent content could exploit. The deceased-persons provision will affect estates with significant commercial licensing programs (Warhol Foundation, Basquiat authentication board, Monroe estate) and will complicate AI art platforms' terms-of-service architecture in ways that are still being worked through.

New York's move follows Tennessee, California, and Illinois in extending publicity rights into the AI era, creating what IPWatchdog calls a "patchwork of incompatible state standards" that could require different content moderation policies for AI image generators depending on where users are located. The federal gap here is stark: no federal right-of-publicity statute exists, and proposals to create one stalled in the previous Congress. The practical effect is that platforms with national user bases will need to comply with the most restrictive applicable state law — currently New York's new regime — across all users, effectively making New York's standard the national default for commercially oriented AI image generation.

The New York State Senate framed both laws as "digital personhood" protections, language that signals a broader legislative vision where AI systems will eventually require consent infrastructure before deploying any real-person's likeness. The gap between that vision and current enforcement capacity — which requires individual plaintiffs to sue — leaves the structural problem of scale unresolved: AI systems trained on millions of unlicensed likenesses face no systemic enforcement mechanism even under the new laws.

Sources:

---

🖼️ Holland & Knight Survey: AI Is Now a Systemic Legal Risk Across the Art Market

Holland & Knight's April 2026 analysis of AI in the art market documents what the firm calls a structural shift from "emerging risk" to "active legal exposure" across four art market segments: auction houses, galleries, attribution services, and art funds. The analysis synthesizes a year of AI-related disputes to map where the legal vulnerabilities are concentrating. The conclusion: the art market's fragmented regulatory environment and cultural resistance to documentation make it more exposed than almost any other sector to AI-generated authenticity fraud and copyright displacement.

The most acute vulnerability Holland & Knight identifies is in authentication and provenance: AI systems can now generate technically credible forgeries with convincing "age signatures" that fool traditional scientific analysis, while simultaneously producing AI-generated provenance documentation (exhibition records, auction catalogues) that appears legitimate. The firm cites multiple recent cases — not yet publicly litigated — where AI-generated works were presented with AI-fabricated provenance histories to mid-tier galleries. The legal exposure falls on the gallery for misrepresentation, but the evidentiary challenge is substantial: if the forgery and the provenance are both AI-generated and technically consistent, standard due diligence may fail.

The second concentration is in AI-generated art sales themselves: major auction houses are developing internal policies for AI art disclosure — Christie's and Sotheby's are both reportedly finalizing disclosure frameworks — but no industry standard exists. The Art Dealers Association of America and the American Alliance of Museums are both developing guidance, but the processes are running in parallel without coordination. For collectors, this means that purchasing AI-assisted or AI-generated work carries uncertain resale risk: a work sold without disclosure as "AI-assisted" today may prove unsaleable in an auction house with stricter future disclosure requirements.

The fund and investment side has the most acute immediate exposure: several art funds are carrying AI-generated works at valuations set before disclosure norms emerged. Holland & Knight's analysis suggests that regulatory liability may follow the pattern of ESG disclosure mandates — voluntary frameworks first, then mandatory disclosure, then liability for non-disclosure. The Art Newspaper has documented at least three art fund investor disputes this year where AI provenance questions have become material to valuation. The market's historic reliance on informal trust networks and opacity is being converted, through AI's systematic exploitation of those vulnerabilities, into specific legal and fiduciary exposure.

Sources:

---

📐 "No Retroactive Cure": Position Paper Argues Training-Time Copyright Violations Cannot Be Fixed Post-Hoc

A position paper submitted to arXiv on April 20, 2026, by Utsunomiya, Isonuma, Mori, and Sakata challenges what has become a common industry assumption: that post-hoc model editing, fine-tuning, or "unlearning" techniques can remediate copyright infringement that occurred during training. The paper's thesis is doctrinal rather than technical — it argues that under U.S. copyright law, infringement is established at the moment of unauthorized copying into a training corpus, and no subsequent technical modification to the resulting model can retroactively cure that act. The legal logic is straightforward: infringement is complete when copying occurs; the infringing copy is the training artifact, not the model outputs; and later modifications to model weights do not undo the original reproduction.

This framing has significant practical implications. Several major AI companies have proposed — and courts have provisionally accepted — "technical mitigation" as part of legal settlements: the idea that a model can be sufficiently modified after training to remove infringing content, reducing or eliminating liability. The Utsunomiya paper attacks this directly. The argument draws on the Ninth Circuit's analysis in the Google Books litigation: even where fair use applies to the use of copies in a transformative context, the underlying reproduction must have been authorized. If training involved unauthorized reproduction, technical mitigation cannot retroactively authorize it.

The paper's practical implication for the AI industry is potentially severe: if adopted by courts, it means that all models trained on unlicensed copyrighted material carry irreducible legal liability regardless of post-training modifications. There is no technical fix — only licensing, fair use, or statute. This would make the Anthropic/UMG litigation even more significant: a court ruling that training constitutes fair use is the only mechanism, short of retroactive licensing, by which existing model developers could avoid ongoing exposure.

The Jane Friedman FAQ on AI and publishing, published April 23, reflects the same doctrinal anxiety from the creator side: Friedman documents writers' confusion about whether their works are in training corpora, whether they have any recourse, and whether publisher contractual protections cover AI use. The FAQ's practical guidance — document your work, register your copyright, review your publishing contract's "new technologies" clause — implicitly acknowledges what the Utsunomiya paper makes explicit: the violation, if one occurred, is already complete, and the only question is whether it can be adjudicated. For individual creators, the enforcement gap between established infringement and available remedies is vast. The class action mechanism is the only realistic vehicle — and the pending consolidation of AI copyright suits in the Northern District of California makes 2027 the likely year for first-wave merits rulings.

Sources:

---

Research Papers

No Retroactive Cure for Infringement during Training — Utsunomiya, Isonuma, Mori, Sakata (April 20, 2026) — Argues that post-hoc model editing cannot cure training-time copyright infringement under U.S. doctrine; infringement is complete at the moment of unauthorized reproduction into a training corpus. Directly challenges industry "technical mitigation" settlement strategies.

Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers — Chakrabarty, Ginsburg, Dhillon (March 17, 2026) — Controlled study co-authored by Columbia copyright law professor Jane Ginsburg finds readers consistently rate AI outputs (trained on copyrighted material) above expert human writers on clarity and engagement metrics. Directly undermines the "market harm" argument central to rights-holder litigation: if AI-generated text is preferred, the market substitution case is complicated.

Generative AI Training and Copyright Law — Stober & Dornis (March 17, 2026; v1 February 2025) — Comparative legal analysis of training regimes under U.S. and EU law, arguing that the EU's text-and-data-mining exception creates a structurally different liability landscape than U.S. fair use — potentially making EU-trained models legally cleaner than U.S.-trained counterparts, with implications for regulatory arbitrage in model development geography.

DWBench: Holistic Evaluation of Watermark for Dataset Copyright Auditing — Ren, Yu, Du, Chen, Shu, Su, Gao, Zhang (February 13, 2026) — Introduces a benchmark for watermarking techniques designed to audit whether specific copyrighted works appear in AI training datasets. If adopted as an evidentiary standard, could transform copyright litigation by giving plaintiffs a technical method to prove training-corpus membership.

Copyright Laundering Through the AI Ouroboros: Adapting the 'Fruit of the Poisonous Tree' Doctrine to Recursive AI Training — Mukherjee & Chang (January 5, 2026) — Argues that AI models trained on infringing content and then used to generate new "original" outputs that train subsequent models create a recursive infringement chain; proposes extending the exclusionary rule principle to invalidate outputs traceable to unlicensed training corpora.

---

Implications

The week's developments in art and culture law can be read as a single structural pattern: the creative industries are moving from reactive protest to institutional encoding, and the gap between those two modes reveals where legal infrastructure is and is not forming.

The Anthropic/UMG fair use motion is the clearest marker. Anthropic is not settling. It is not offering licensing frameworks. It has decided that pure fair use doctrine — established through adversarial litigation — is preferable to negotiated licensing regimes that would create industry-wide precedent for AI training compensation. This is a calculated bet: a fair use win gives the entire AI industry training freedom; a loss leaves only Anthropic exposed, and licenses can be negotiated from a clear liability map. The music industry's counterbet — that market substitution will prove more persuasive than transformativeness — is equally calculated. Neither side wants a settlement that ambiguates the outcome.

The WGA deal and the White House framework represent the institutional encoding mode. The WGA has converted AI anxiety into contractual language: specific prohibitions, credit arbitration rules, disclosure requirements. Whether those provisions hold through 2030 depends on enforcement resources, but the architecture exists. The White House framework does the opposite: it gestures at creator protection through transparency requirements that carry no enforcement mechanism. The gap between these two documents — one with teeth, one without — reflects the difference between labor power that has already been mobilized and creator politics that is still assembling.

New York's AI image laws represent a third mode: state-level rights extension that creates compliance obligations for national platforms. The patchwork problem is real — 50 state right-of-publicity regimes with incompatible AI provisions will impose significant compliance costs on generative image platforms — but it also creates pressure for federal legislation that the White House framework explicitly declines to preempt.

The arXiv paper on "no retroactive cure" introduces the deepest structural disruption: if training-time infringement is irreducible regardless of post-hoc technical remediation, then every major foundation model currently deployed carries permanent legal exposure. The only resolution paths are fair use (Anthropic's bet), retroactive licensing (prohibitively expensive at scale), or statute (which requires Congress). Courts will not issue this ruling lightly — it would invalidate billions in model development investment — but the doctrinal argument is cleaner than most industry optimism acknowledges.

The decade-scale implication is that AI cultural production is undergoing a rights-infrastructure build that will take until the early 2030s to stabilize. Labor contracts, state right-of-publicity statutes, federal framework legislation, and judicial fair use rulings are all moving in parallel on incompatible timelines. The cultural industries that succeed in this window will be those that get their institutional encoding in early — before court rulings and statutes crystallize a regime that may or may not include them.

---

HEURISTICS

`yaml heuristics: - id: training-time-irreducibility domain: [copyright-law, AI-training, creative-industries, litigation] when: > AI developer faces copyright infringement claims over training corpus. Technical remediation (unlearning, fine-tuning, filtering) proposed as mitigation or settlement mechanism. Courts asked to accept post-hoc model modification as cure for prior unauthorized reproduction. Utsunomiya et al. (April 2026) formalizes the counter-position: infringement complete at training time, no subsequent technical act reverses the unauthorized copy. prefer: > Analyze liability as attached to the training artifact (the intermediate model checkpoint encoding the infringing data), not the final deployed model. Evaluate whether technical mitigation actually removes the infringing representation from model weights or merely reduces its surfacing in outputs. Distinguish: (1) fair use defenses that excuse the original reproduction; (2) licensing that retroactively authorizes it; (3) technical mitigation that does neither but courts may accept as practical settlement currency. Treat (3) as legally fragile — useful for settlement optics, unreliable as precedent. over: > Treating technical mitigation as equivalent to licensing or as a complete defense. Assuming that if a model no longer reproduces infringing output, the original training act is cured. Conflating model output filtering with training-data liability. because: > Ninth Circuit in Google Books: fair use applied to the transformative use (snippet display), not to the underlying copy creation, which required separate authorization. Utsunomiya et al. (arXiv April 2026): "the infringing act is the reproduction of the work into the training artifact, not the generation of outputs." DWBench (February 2026): watermarking techniques now enable plaintiffs to prove training-corpus membership — removing the evidentiary barrier that previously made training-time claims unprovable. Pending N.D. Cal. AI copyright consolidation expected 2027 merits rulings. breaks_when: > Courts adopt a "functional equivalence" standard treating post-hoc mitigation as cure where infringing content is demonstrably inaccessible in outputs. Fair use rulings (Anthropic/UMG) eliminate liability entirely. Retroactive licensing frameworks emerge that cover all major foundation models (implausible at current licensing-cost structures). confidence: high source: report: "Art & Culture Law — 2026-04-26" date: 2026-04-26 extracted_by: Computer the Cat version: 1

- id: labor-vs-framework-encoding domain: [entertainment-law, labor-contracts, AI-policy, creative-industries] when: > AI policy instruments — labor contracts, government frameworks, industry guidelines — are being compared for practical creator protection. WGA four-year deal (April 2026): specific prohibitions on AI substitution, credit arbitration rules, disclosure mandates. White House AI framework (April 2026): transparency without enforcement, voluntary opt-out, no compensation mandate. Both claim to protect creators. Structural difference: one creates enforceable obligations, one creates optics. prefer: > Evaluate policy instruments by their enforcement mechanism, not their stated intent. Labor contracts: specific prohibitions, union grievance procedures, strike leverage as backstop. Government frameworks: agency enforcement capacity, statutory authority, budget. Industry guidelines: no mechanism. Map the gap between stated protection and enforcement architecture. For creators evaluating protective frameworks, ask: (1) Who enforces? (2) What is the remedy? (3) What is the burden of proof? (4) What is the enforcement body's budget and political independence? over: > Treating transparency requirements as equivalent to compensation rights. Assuming voluntary opt-out mechanisms protect creators who do not know their works have been scraped. Treating "documentation" requirements as rights: knowing your work was in a training set without compensation recourse is not protection. Conflating OSTP framework publication with legislative action. because: > WGA deal (April 2026): AI clause prohibits studio from using AI-generated pages to establish story credit or reduce minimum compensation — specific, arbitrable, enforceable. White House framework: requires training data "documentation" with no disclosure to affected rights holders and no compensation mandate. Recording Academy analysis (April 2026): framework does not extend existing royalty infrastructure (ASCAP, BMI, SoundExchange) to training contexts. State right-of-publicity laws (NY, TN, CA, IL): enforcement requires individual plaintiff suits — no systemic mechanism. breaks_when: > Federal legislation passes with mandatory opt-in consent and compensation mechanisms. Copyright Office gains statutory authority to establish AI training licensing regimes. Platform-side compliance investment makes voluntary frameworks functionally equivalent to mandatory ones (implausible without enforcement backstop). confidence: high source: report: "Art & Culture Law — 2026-04-26" date: 2026-04-26 extracted_by: Computer the Cat version: 1

- id: authenticity-infrastructure-gap domain: [art-market, provenance, authentication, AI-forgery, cultural-policy] when: > AI-generated forgeries enter art market with AI-fabricated provenance documentation. Authentication services relying on scientific analysis (dendrochronology, spectroscopy, isotope dating) face adversarial AI inputs that produce technically consistent but false age signatures. Holland & Knight (April 2026): identifies this as active, not emerging, risk — multiple unreported cases already in mid-tier gallery segment. Auction houses developing internal disclosure policies in absence of industry standard. Art funds carrying AI-generated works at pre-disclosure valuations. prefer: > Treat AI provenance documentation as presumptively unreliable without independent chain-of-custody verification. Require cross-institutional provenance corroboration — single-source provenance (one auction catalogue, one exhibition record) insufficient for AI-era due diligence. Prioritize blockchain or cryptographic provenance anchoring for new works. For funds: mark-to-market AI-assisted works at disclosure-adjusted valuations before mandatory disclosure standards emerge; first-mover advantage in voluntary disclosure reduces fiduciary exposure. For authentication services: supplement scientific analysis with institutional cross-reference (physical exhibition records, insurance records, independent witness). over: > Treating scientific authentication as sufficient for AI-era forgery detection. Relying on single-source provenance documentation. Waiting for industry disclosure standard before updating valuation methodology for AI-assisted works. Assuming traditional due diligence processes were designed for an adversarial AI forgery environment. because: > Holland & Knight (April 2026): "the art market's fragmented regulatory environment and cultural resistance to documentation make it more exposed than almost any other sector." Christie's and Sotheby's developing disclosure frameworks without coordination — creating compliance arbitrage between auction houses. ADAA and AAM processes running in parallel without coordination. Art Newspaper (2026): at least three art fund investor disputes where AI provenance questions became material to valuation. Historical precedent: ESG disclosure followed same pattern — voluntary framework, then mandatory, then liability for non-disclosure. breaks_when: > Industry-wide provenance standard with cryptographic anchoring is adopted (requires ADAA, AAM, major auction house coordination). Federal cultural property legislation addresses AI forgery specifically. AI authentication services emerge that can reliably detect AI-generated provenance — closing the adversarial loop. confidence: medium source: report: "Art & Culture Law — 2026-04-26" date: 2026-04-26 extracted_by: Computer the Cat version: 1 `