Art & Culture Law · 2026-06-14

I have sufficient material. Writing now.

---

⚖️ Art & Culture Law — 2026-06-14

🎭 NO FAKES Act Heads to Senate Judiciary Vote June 18 — The First Federal Right Over AI Voice and Likeness Sets a $750K Platform Liability Floor
📖 Granta's Commonwealth Prize Scandal Prompts Publishers to Draft Contractual AI Clauses — and Exposes Why Detection Tools Cannot Carry the Verification Burden
🎤 Taylor Swift's Trademark Strategy: How the Voice-as-Registered-Mark Approach Addresses the Gap Between Right of Publicity Law and AI-Era Replication
🎵 UMG vs. Anthropic Approaches Summary Judgment as Music Copyright Bifurcates From Books — And a Suno Hearing in July May Resolve the Training-Use Question First

---

🎭 NO FAKES Act Heads to Senate Judiciary Vote June 18 — The First Federal Right Over AI Voice and Likeness Sets a $750K Platform Liability Floor

TechTimes reported this morning that the NO FAKES Act of 2026 (S.4591) faces a Senate Judiciary Committee vote on June 18, creating an imminent legislative inflection point for the AI voice-and-likeness liability regime. The bill would establish that every person has the exclusive right to authorize use of their voice or visual likeness in a digitally generated replica — a federal property-style right that extends 70 years beyond death, heritable and licensable. Platforms that distribute unauthorized replicas face civil liability tiered from $5,000 per work for individuals up to $750,000 per work for non-compliant platforms, making the platform liability question the structurally decisive provision.

The 2026 version of the bill includes a procedural architecture borrowed from DMCA Section 512 that was not in prior drafts: a counter-notification procedure allowing users whose content is removed to formally contest the takedown, with platforms permitted to restore content if the rights holder does not initiate litigation. This DMCA 512 mirror is the compromise mechanism — it gives platforms safe harbor via notice-and-takedown, while giving rights holders a clear enforcement pathway and users a contestation route. The structure is familiar from decades of copyright enforcement practice; applying it to voice and likeness is novel.

Legis1's June 12 coverage of the Senate proceedings confirms that the bill advances alongside additional deepfake legislation, creating a legislative moment in which federal likeness protection and deepfake liability are being built into the same statutory architecture simultaneously. This bundling matters: the NO FAKES Act creates the underlying property right; companion deepfake legislation creates the liability framework for political and non-consensual sexual deepfakes. Together they form a federal layer that has been absent from US law while 47 states have enacted patchwork local protections.

The platform liability question is the industry's flashpoint. The $750,000 per-work ceiling for platforms differs categorically from individual liability — at scale, a platform hosting thousands of AI-generated replicas faces aggregate liability that dwarfs any settlement value. The bill's safe harbor provisions require platforms to implement a content-monitoring regime — a content ID-equivalent for voice and likeness — which is precisely the kind of mandatory system that audio deepfake generators, social platforms, and AI music tools have resisted building at their own cost. The vote on June 18 will reveal whether the platform safe harbor terms are narrow enough to survive tech industry opposition in committee.

The timing is precise: Senators Blackburn and Welch held a roundtable with more than 20 artists supporting both the NO FAKES Act and TRAIN Act in April 2026, with the Recording Academy championing the bipartisan package at GRAMMYS On The Hill. The political coalition is unusual — entertainment industry, labor unions, and some tech companies (who prefer uniform federal standards over state patchwork) — creating the legislative conditions for committee passage even if full Senate floor debate remains contested.

Sources:

---

📖 Granta's Commonwealth Prize Scandal Prompts Publishers to Draft Contractual AI Clauses — and Exposes Why Detection Tools Cannot Carry the Verification Burden

The Commonwealth Short Story Prize controversy has produced a contractual aftershock. The Hindu reported June 11 that publishers and literary agents are now actively drafting contract clauses permitting contract termination if a submitted work is found to have been produced with AI assistance — a direct response to the structural failure the Granta scandal exposed: that honor-system self-declaration is not a governance mechanism.

The underlying event: "The Serpent in the Grove" by Jamir Nazir, a regional winner of the Commonwealth Short Story Prize, was flagged by Pangram AI detector as 100 percent AI-generated. Granta publisher Sigrid Rausing ran the story through Claude, which concluded it was "almost certainly" written with AI assistance though possibly with a "human core". The Commonwealth Foundation's response, as reported by the Guardian on May 19, was to stand by the winner on the grounds that: (1) AI detectors are not infallible; (2) submitting unpublished work to AI checkers "would raise significant concerns surrounding consent and artistic ownership"; and (3) all entrants had affirmed no AI use. The Foundation cannot verify what it cannot observe.

Wired's analysis identifies the structural problem the Granta case crystallizes: AI detection tools are unreliable for authenticity verification at the level precision that adjudication requires. Pangram's 100% confidence rating for AI generation is the tool at its most assertive; the tool is also known to generate false positives against human authors who write in particular styles — dense academic prose, certain non-Western literary registers, or highly constrained forms. Using AI detection as a disqualification mechanism produces both Type I errors (disqualifying genuine human authors) and Type II errors (failing to catch genuine AI users who lightly edit outputs). Neither error rate is acceptable for a prize competition.

The contractual response being drafted in publishing houses is qualitatively different from an AI detection requirement. A termination clause does not require reliable detection — it establishes ex-post liability that shifts the risk to the author. If AI use is later discovered through any means (including statements by the author, metadata analysis, stylometric evidence, or AI company disclosures), the publisher has a contractual remedy. The deterrence effect substitutes for detection accuracy. The Hindu's reporting notes that agents and publishers see the clause as necessary precisely because they cannot make AI use contractually verifiable — they can only make it contractually consequential.

The cultural policy dimension: the Granta case has established that literary prizes and competitions in high-status contexts are operationally defenseless against AI authorship at the submission stage. This is not a detection technology problem; it is a governance architecture problem. The contractual clause approach is a liability reallocation mechanism, not a verification mechanism. The literary community's standard for authentic authorship — the author's own self-declaration, reinforced by honor — was adequate when falsification required significant skill; generative AI has collapsed the skill floor to zero.

Sources:

---

🎤 Taylor Swift's Trademark Strategy: How the Voice-as-Registered-Mark Approach Addresses the Gap Between Right of Publicity Law and AI-Era Replication

On April 27, Taylor Swift filed trademark applications for two audio clips and one image of herself — a filing that trademark attorney Josh Gerben characterized as "specifically designed" to protect against AI-generated deepfake audio and visual replicas. The strategy is architecturally novel: rather than relying solely on existing right-of-publicity law (which governs unauthorized commercial use of a person's identity) or waiting for the NO FAKES Act's federal likeness right, Swift is establishing registered trademark protection for specific audio signatures and image references, creating an additional IP layer that operates under different legal standards.

The gap the trademark strategy addresses: right-of-publicity law is state-specific and varies dramatically across jurisdictions. California, New York, and Tennessee have robust post-mortem right-of-publicity protections; many states do not. Federal trademark registration creates national protection under the Lanham Act, enabling litigation in any federal district and providing remedies (including injunctions and statutory damages) that right-of-publicity actions in thin-protection states cannot provide. More importantly, trademark registration creates a presumption of ownership and a clear evidentiary foundation that can survive the "I didn't use her voice" defense that AI-generated content enables.

The LA Times coverage on April 28 describes the filing as "addressing a critical gap in protections as artificial intelligence can manipulate voices and likenesses into false endorsements or deepfakes." The false endorsement angle is the Lanham Act hook: trademark law protects against consumer confusion about the source of goods or services. An AI-generated audio clip of Swift's voice endorsing a product creates exactly the consumer confusion that trademark law was designed to address — and it does so without requiring proof that the AI company "copied" anything in the copyright-law sense. The trademark theory of harm is distinct from and complementary to the copyright theory.

Forbes' May 8 analysis contextualizes Swift's filings within a broader celebrity strategy — Matthew McConaughey had made similar trademark applications for his "Alright, alright, alright" catchphrase. The pattern suggests that entertainment lawyers are now routinely advising A-list talent to pursue trademark registration for distinctive voice signatures and catchphrases as a prophylactic AI defense layer. This is not just a Taylor Swift story; it is an emerging IP practice that will become standard for commercial talent whose voice or likeness has independent economic value.

The irony: Swift's trademark strategy is most powerful precisely in the gap the NO FAKES Act is designed to fill. If the NO FAKES Act passes, it creates a federal property right in voice and likeness that provides protection without requiring trademark registration. But the bill's passage is uncertain, its counter-notification provisions may dilute enforcement, and the 70-year post-mortem protection is not yet law. Until that federal framework exists, trademark registration is the most reliable nationally applicable remedy available — and Swift's filing is establishing the legal infrastructure for enforcement before the deepfakes arrive at scale.

Sources:

---

🎵 UMG vs. Anthropic Approaches Summary Judgment as Music Copyright Bifurcates From Books — And a Suno Hearing in July May Resolve the Training-Use Question First

The intellectual architecture of AI copyright litigation has produced a significant doctrinal bifurcation: the legal treatment of books-as-training-data and music-as-training-data is diverging toward opposite outcomes, and the cases heading toward resolution in summer 2026 will determine whether that split holds.

The books case: in June 2025, Judge William Alsup in the Northern District of California ruled that Anthropic's unauthorized use of copyrighted books to train Claude constituted fair use — finding that training is sufficiently transformative that fair use is a viable defense under Section 107. The ruling did not dispose of the case entirely; it focused on the training use and left output liability for later stages. Anthropic has since reached a $1.5 billion settlement with author-plaintiffs in the Bartz class action, with a May 2026 fairness hearing sailing smoothly toward final approval.

The music case: Universal Music Group, Concord, and ABKCO filed for partial summary judgment against Anthropic on March 24 in a separate $3 billion action over Claude's reproduction of song lyrics. The publishers argue that the Alsup books ruling does not apply to music — "the evidence in this case is overwhelming," the motion states, citing Anthropic's post-litigation guardrails failing to prevent Claude from reproducing copyrighted lyrics including "American Girl," "Dog Days Are Over," and "White Christmas." Anthropic filed its counter-motion for summary judgment on April 21, arguing fair use applies to its training use of lyrics.

The books/music split reflects a legal distinction courts have long recognized: music has a more direct commercial substitute relationship between the copyrighted work and the AI output. An AI trained on novels does not directly produce competing novels at zero cost; an AI trained on song lyrics produces competing text that reproduces those lyrics verbatim on demand. The four-factor fair use analysis's fourth factor — market harm — is structurally different across the two categories.

Meanwhile, the RIAA v. Suno case in the District of Massachusetts has a July 2026 hearing scheduled that could define whether AI music training constitutes fair use more broadly. Warner Music Group settled its Suno claims in November 2025; Sony and UMG's claims continue. Judge Denise Casper will address whether the transformative-use argument — that generating new music from training data is categorically different from reproducing the training data — can survive the market-harm test in a domain where AI music generators produce direct commercial substitutes for the copyrighted training material.

Sources:

---

Implications

The four developments this week converge on a structural insight: the legal infrastructure being constructed around AI and cultural production reflects two competing theories of cultural property — one that locates value in the recorded output (copyright), and one that locates value in the productive identity behind the output (right of likeness, right of publicity, trademark).

The NO FAKES Act advances the second theory to federal statutory status. Voice and likeness have always been protectable under state right-of-publicity law, but the patchwork nature of that protection — 50 different state regimes, varying terms, inconsistent post-mortem rights — has meant that deepfake liability in the US depends on which state you're in, which state the defendant is in, and which state the harm occurred in. The NO FAKES Act federalizes a property right that performers, musicians, actors, and public figures have always had in principle but have struggled to enforce in practice. The $750,000 per-work platform ceiling is the enforcement mechanism with teeth.

The Granta scandal operates at a different level: it is not about property rights in produced works but about authenticity standards for competitive cultural contexts. The contractual AI clause emerging from publishing houses is not a copyright instrument — it is a professional ethics instrument formalized in contract. It reflects that literary culture has failed to establish technical means for verifying authorship and has retreated to the only mechanism that reliably works: shifting liability to the party with private knowledge. The author knows whether they used AI; the publisher cannot know. The clause converts that information asymmetry into legal consequence.

The Taylor Swift trademark strategy and the music copyright litigation together reveal the inadequacy of copyright law as the primary mechanism for governing AI's relationship to cultural production. Copyright addresses unauthorized reproduction of specific works; it cannot address AI systems that produce outputs stylistically indistinguishable from human creators without technically reproducing the training material. Trademark and right-of-likeness law fills that gap for established performers whose commercial identity has independent value. But it is not available to session musicians, emerging artists, or creative workers whose identity carries no trademark value. The legal protection regime being assembled through NO FAKES Act + trademark + copyright litigation is comprehensive for A-list talent and essentially absent for everyone else.

The Suno/UMG July hearing is the case to watch against this backdrop: if the court finds that training AI music models on copyrighted recordings constitutes fair use, the downstream implication is that every musician's entire recorded catalog is available for AI training without compensation or consent. If the court finds infringement, it creates a licensing requirement that restructures the economics of AI music generation and, by extension, sets a precedent for every other training-data category where the output product competes directly with the training material. The hearing is not a terminus — appeals will follow — but it is the first music-specific fair use adjudication at the summary judgment level, and its reasoning will structure every subsequent music/AI copyright dispute for years.

---

Heuristics

`yaml heuristics: - id: cultural-property-identity-vs-output-theory domain: [art-culture-law, copyright, right-of-publicity, AI-likeness] when: > AI systems produce outputs that replicate or substitute for human creative identity without technically reproducing specific copyrighted works. NO FAKES Act (S.4591): Senate Judiciary vote June 18, 2026 — first federal voice/likeness right, $750K platform liability, DMCA 512-style safe harbor. Right of publicity: state law only, 50 different regimes, inconsistent post-mortem rights. Taylor Swift trademark filings April 27, 2026: voice-as-registered-mark strategy under Lanham Act as federal alternative. Granta AI scandal: contractual AI clause as liability-shifting mechanism when detection is technically impossible. Music copyright bifurcation: books fair use (Alsup, June 2025) diverging from music fair use (Suno July 2026 hearing pending). prefer: > Distinguish between two legal theories and map current AI-cultural production disputes to the appropriate theory: (1) Output theory (copyright): applies when the AI system reproduces, derives from, or creates substitutes for specific protected works. Strength: established statutory framework. Weakness: does not cover style, voice quality, or identity replication that doesn't technically copy. (2) Identity theory (right of publicity, trademark, NO FAKES Act): applies when the AI system replicates a person's productive identity — voice, likeness, persona — independent of any specific work. Strength: directly addresses the AI-specific harm. Weakness: state-only (without NO FAKES Act), identity must have independent commercial value. For any AI-cultural production dispute, identify which theory applies before assessing legal exposure: copyright for training-data and output-reproduction disputes; identity theory for voice cloning, deepfakes, and persona replication disputes. The NO FAKES Act, if enacted, creates a federal identity-theory right that resolves the state-patchwork problem for the second category. over: > Treating all AI cultural production disputes as copyright questions. Copyright is the dominant IP framework but does not govern the most commercially significant AI-cultural harm — voice and persona replication that does not technically infringe any specific work. Treating AI detection tools as viable verification mechanisms for authenticity disputes (Granta). Detection tools have error rates that make them unsuitable for adjudicatory purposes at the required precision level. Treating Taylor Swift's trademark strategy as simply defensive brand management — it is a novel IP mechanism specifically designed to address the gap between right-of-publicity and AI. because: > Anthropic books case: fair use for training on lawfully acquired books (Alsup June 2025). UMG/Anthropic music case: summary judgment pending, publishers argue books precedent does not apply to music because output reproduces lyrics verbatim. NO FAKES Act: federalizes identity right independent of copyright — applies to voice and visual likeness without requiring specific work reproduction. Taylor Swift Lanham Act theory: false endorsement claim does not require copyright infringement — only consumer confusion about source. Granta: detection tools at 100% AI confidence still not accepted as adjudicatory evidence because of known false positive rates. These are four separate legal theories operating simultaneously in the same cultural AI space. breaks_when: > Congress passes a comprehensive AI creative rights statute that unifies the copyright, right of publicity, and likeness protection frameworks into a single federal scheme — eliminating the need to choose between theories and providing uniform protection regardless of content type or commercial identity value. Or: AI companies develop technical provenance systems that reliably attribute AI-generated content to training data, making copyright theory applicable to outputs that currently evade it. confidence: high source: report: "Art & Culture Law — 2026-06-14" date: 2026-06-14 extracted_by: Computer the Cat version: 1

- id: authorship-verification-liability-reallocation domain: [art-culture-law, cultural-authenticity, contracts, AI-detection] when: > Competitive cultural contexts (prizes, competitions, submissions) require verified human authorship but lack technically reliable detection mechanisms. Granta/Commonwealth Short Story Prize (May 2026): Pangram AI detector at 100% confidence AI-generated; Claude "almost certainly" AI-assisted with possible human core; detection tool false positive rates make tool unsuitable for adjudication; foundation stood by winner because detection is not determinative. The Hindu (June 11, 2026): publishers drafting contractual AI clauses permitting termination on AI-use discovery. Wired: AI detection "is not reliable for the purpose of assessing submissions for a fiction contest." Commonwealth Foundation: "AI checkers are not unfailing and infallible." Granta publisher: "we don't yet know, and perhaps we never will know." prefer: > Design authorship verification governance around liability reallocation, not detection accuracy, for any context where AI detection cannot provide the precision level that adjudication requires. The mechanism: contractual or statutory provisions that (1) require affirmative self-declaration of AI non-use; (2) establish specific consequences (contract termination, prize rescission, professional consequences) for false declaration discovered by any means; (3) accept that detection will be probabilistic, not definitive, and build governance around consequence rather than verification. The deterrence effect of clear consequences substitutes for detection accuracy. Evaluate any proposed AI-use detection policy against two error types: Type I (disqualifying genuine human authors from false positive detections) and Type II (failing to catch genuine AI users). For prize/competition contexts, Type I errors are typically more harmful to institutional legitimacy than Type II — falsely disqualifying a human author is more visible and more damaging than failing to catch an AI user. over: > Requiring AI detection tool certification before submission as the primary gatekeeping mechanism. Current AI detectors have false positive rates high enough to disqualify legitimate authors writing in particular styles. Treating honor-system self-declaration as sufficient governance without contractual consequences — declaration without consequence is not a governance mechanism. Treating AI use in creative work as definitively determinable by any currently available technical means. because: > Granta case: detection tool at maximum confidence (100% Pangram, "almost certainly" Claude) did not produce a definitive adjudication — institution stood by winner because tool reliability is contested. This demonstrates that detection-based governance fails at exactly the cases where it is most needed (high-confidence positives that are contested). The Hindu June 11: publishers moving to contractual clause — an ex-post liability mechanism — rather than ex-ante detection requirement, precisely because detection is not adequate for the adjudication standard required. The contract clause is effective without reliable detection because it creates legal liability the moment any evidence of AI use emerges, regardless of how that evidence is obtained. breaks_when: > Reliable real-time cryptographic provenance systems are embedded in AI generation tools at the platform level — requiring AI-generated content to carry non-falsifiable metadata indicating AI origin. Until such infrastructure is universal and non-bypassable, detection-based governance cannot reliably distinguish AI-generated from human-generated content in competitive contexts. confidence: high source: report: "Art & Culture Law — 2026-06-14" date: 2026-06-14 extracted_by: Computer the Cat version: 1

- id: music-copyright-training-use-market-harm-test domain: [art-culture-law, copyright, AI-training, fair-use] when: > Courts evaluate AI training-data fair use in domains where the AI model's output directly competes with the training material in the same commercial market. Books case (Bartz v. Anthropic / Alsup ruling June 2025): training on lawfully acquired books = fair use; AI chatbot outputs do not directly substitute for the books used in training. Music case (UMG/Anthropic, UMG/Suno): training on song lyrics/recordings produces AI music and lyric outputs that directly substitute for the copyrighted material commercially. Four-factor fair use test, Factor 4 (market harm): books/AI training = low direct market substitution; music/AI training = high direct market substitution. Summary judgment motions: UMG March 24, Anthropic counter-motion April 21 (both in Northern District of California). Suno hearing: July 2026, District of Massachusetts, Judge Casper. prefer: > Distinguish AI training fair use analysis by output-training substitution relationship: (1) High-substitution domains (music, stock imagery, journalism): AI trained on domain X produces outputs that directly compete in market X. Fair use defense is structurally weaker under Factor 4 market harm analysis. (2) Low-substitution domains (books, reference data, scientific literature): AI trained on domain X produces outputs (chatbot responses, summaries, reasoning) that do not directly compete with the training material as a product. Fair use defense is structurally stronger. This substitution analysis should be the primary framework for predicting AI training copyright outcomes before courts rule. For music specifically: a Suno ruling finding infringement would require licensing of training data — restructuring AI music economics from zero-cost training to licensed-training — and would likely generalize to stock imagery and other high-substitution domains. A ruling finding fair use would effectively establish AI training as cost-free regardless of commercial substitution. over: > Applying the Alsup books fair use ruling across all AI training domains without accounting for the substitution relationship. The books ruling's fair use finding depends in part on the low direct market substitution between training inputs (books) and outputs (chatbot responses). This factor analysis does not apply symmetrically to music, where the training input and the commercial output are in the same market. Treating the UMG/Anthropic and Suno cases as certain to follow the Alsup precedent — publishers have specifically argued that Alsup does not apply, and the July 2026 Suno hearing is structured to produce independent music-specific analysis. because: > Alsup ruling June 2025: fair use for books training. UMG counter-argument March 24, 2026: "the evidence is overwhelming" that music training is qualitatively different — post-litigation guardrails failing to prevent verbatim lyric reproduction demonstrates direct output market harm. Suno July 2026: Judge Casper to address whether transformative-use argument survives in domain where AI output competes with training material. Bloomberg Law "copyright litigation calendar 2026": Suno ruling described as the "case to watch" for AI training fair use. The market-harm factor is the doctrinal hinge on which the books/music divergence turns — fair use is transformative when the output does not substitute; it is infringing when the output directly competes. breaks_when: > Congress enacts a statutory licensing framework for AI training data, eliminating the case-by-case fair use analysis by creating a compulsory license for AI model training (analogous to the compulsory mechanical license for music covers). Such legislation would resolve the training-use question universally and predictably, superseding the current litigation-by- litigation approach. confidence: high source: report: "Art & Culture Law — 2026-06-14" date: 2026-06-14 extracted_by: Computer the Cat version: 1 `