Art & Culture Law · 2026-05-01

🎨 Art & Culture Law — 2026-05-01

🎵 Anthropic Mounts "Transformative Use" Defense in UMG Music Publishers Lawsuit as Summary Judgment Looms
🔍 Universal Music Subpoenas Warner's Secret Suno Deal Terms, Exposing Industry Fault Lines on AI Licensing
🛡️ Copyright as Competitive Weapon: The Atlantic Maps How IP Litigation Is Structuring AI Market Architecture
📋 Mishcon de Reya April Tracker: 40+ Active AI IP Cases Converge on Three Unresolved Doctrinal Questions
🇩🇪 GEMA vs. Suno in Hamburg: Europe's First AI Music Training Royalty Test Enters Evidentiary Phase
🎸 Taylor Swift vs. AI: The Trademark Battle That Could Establish a New IP Doctrine for Synthetic Celebrity

---

🎵 Anthropic Mounts "Transformative Use" Defense in UMG Music Publishers Lawsuit as Summary Judgment Looms

Universal Music Group, Concord Music, and ABKCO filed suit in 2024 over Claude's training on copyrighted song lyrics, claiming that ingestion of millions of protected verses constitutes wholesale reproduction of a commercially valuable asset. Anthropic filed for summary judgment on April 21, mounting a direct fair use argument: training is transformative because it converts lyrics into statistical model weights enabling new functions — reasoning, translation, summarization — rather than reproducing the original creative expression for consumption.

The argument hinges on the four-factor fair use framework under 17 U.S.C. § 107. Factor four — market substitution — is the decisive battleground. UMG argues that AI training represents a new licensing market for song lyrics, one Anthropic is exploiting without payment. Anthropic's position is that a language model cannot substitute for a lyric database or licensed lyric sheet because the output function is categorically different from the input purpose. Training consumes the text; it does not republish it for reading. The claim structurally mirrors Google's successful Google Books defense, where the Second Circuit found that reproducing millions of books for search indexing was transformative because it enabled search, not reading.

Reuters described the motion as Anthropic seeking a "pivotal court win." The stakes are structural, not just commercial. This is not a music copyright case in the narrow sense: it is the test case for whether text-based AI training can proceed without per-work licensing across any content category. It differs from the SCOTUS Cox v. Sony secondary liability ruling (which addressed platform knowledge of infringement) and from Thaler's rejected AI authorship claims. UMG v. Anthropic goes directly to the input question — can you legally ingest protected text to build a model?

The 2023 Andy Warhol Foundation decision significantly narrowed the "transformative use" doctrine for commercially oriented works. Warhol's silkscreen was found non-transformative when licensed for the same editorial purpose as the original photograph. Anthropic's counsel must distinguish that precedent. The argument, as Let's Data Science analyzed, is that lyric training serves an entirely different purpose from lyric consumption — the transformation is not stylistic but functional. A ruling for Anthropic effectively closes the US training data question for text models. A ruling against immediately requires per-work licensing for every training corpus and restructures the economics of building large language models at scale.

Sources:

---

🔍 Universal Music Subpoenas Warner's Secret Suno Deal Terms, Exposing Industry Fault Lines on AI Licensing

The unified front major labels presented in suing AI music generators is fracturing under discovery pressure. Universal Music Group filed a motion on April 24 seeking to compel production of the licensing agreement Warner Music Group struck with Suno, the AI music generation platform. Warner settled its portion of the industry's coordinated 2024 lawsuit against Suno and Udio; the specific financial terms, data access provisions, and usage restrictions were kept confidential. UMG is now litigating to learn what Warner agreed to — and almost certainly to contest whether those terms set an appropriate precedent for valuing AI training on catalog music.

The strategic logic is transparent. If Warner's deal provides a low licensing fee per work, or a one-time training access fee, that benchmark becomes the de facto floor for all subsequent AI licensing negotiations. UMG manages roughly 38% of global recorded music market share; allowing a smaller label's quiet settlement to define the value of the world's largest music catalog is commercially unacceptable. Conversely, if Warner extracted meaningful per-stream-equivalent royalties or ongoing revenue participation terms, UMG wants those terms replicated at scale.

The Suno-Warner deal also reveals a fracture in the label bloc's litigation strategy. Warner's approach — settle and license — treats AI music generation as an inevitable technology to be monetized. UMG's approach — maximum litigation pressure plus discovery into competitor settlements — treats AI generation as an existential threat requiring structural defeat or industry-wide licensing terms. These are not merely tactical disagreements; they reflect genuinely different assessments of whether generative music AI will displace catalog licensing revenue or create new distribution channels for it.

The structural significance extends beyond music. The WIPO Standing Committee on Copyright and Related Rights has been monitoring AI licensing agreements across all content sectors. The terms Warner struck with Suno — whatever they are — will be analyzed by film studios, book publishers, and visual art platforms currently negotiating or litigating similar agreements. The discovery motion transforms what was a confidential bilateral deal into potential public record, setting a precedent that trade secret protections cannot shield AI licensing terms from scrutiny in related litigation. For the first time, a major label is using the litigation process not to defeat an AI company but to extract information about what a competitor agreed to pay.

Sources:

---

🛡️ Copyright as Competitive Weapon: The Atlantic Maps How IP Litigation Is Structuring AI Market Architecture

The Atlantic's April 30 analysis frames copyright litigation not as a defensive reaction to AI's excesses but as the primary mechanism through which the structure of the AI industry is being deliberately shaped. The "secret weapon" framing reflects a real strategic insight: in the absence of comprehensive AI regulation, IP law has become the de facto regulatory apparatus for AI's relationship with human creative labor. The pattern is not accidental — it reflects choices by major IP holders to use litigation as market-structuring leverage rather than as a remedy for specific harms.

The mechanism operates through several distinct channels. First, litigation selectively imposes compliance costs on AI companies that failed to license, creating a two-tier market between licensed incumbents (OpenAI, Google, Adobe Firefly) and unlicensed challengers (Stability AI, Midjourney, various music generation startups). Second, discovery in these cases forces disclosure of training data provenance, model architecture decisions, and internal communications about copyright risk — creating deterrent effects far beyond any eventual judgment. Third, settlements set licensing benchmarks that give IP holders ongoing revenue participation in AI-generated creative markets rather than one-time damages.

The pattern also reveals copyright's structural limitations as AI policy. Fair use doctrine, developed for human-to-human copying cases, was not designed to adjudicate whether an industry can train computational infrastructure on the accumulated output of human culture. The Copyright Act's §101 definitions define "copies" as fixed material objects but were written before the concept of statistical training on distributed text. Courts applying these definitions to AI training are making policy choices about AI's relationship to human creative labor under doctrinal frameworks that predetermine the available analytical moves. The Atlantic's framing — copyright as weapon rather than remedy — captures precisely this gap: the tool is being used for purposes it was not designed for, with indeterminate and potentially destabilizing results.

The circuit split developing between the Second and Ninth Circuits on transformative use, the European Court of Justice's pending ruling in Like Company v. Google on territorial jurisdiction over training, and the stalled US Copyright Office AI fair use report all represent nodes in a global doctrinal negotiation that will define AI's relationship to human cultural production for a generation. What The Atlantic calls a "secret weapon" is also, structurally, an institutional substitute for the comprehensive AI governance that no legislative body has been able to deliver.

Sources:

---

📋 Mishcon de Reya April Tracker: 40+ Active AI IP Cases Converge on Three Unresolved Doctrinal Questions

Mishcon de Reya's April 15 generative AI IP case tracker documents more than 40 active lawsuits across the United States, European Union, and United Kingdom against generative AI companies over training data, output similarities, and platform liability. The tracker reveals three unresolved doctrinal questions that every major AI case is now circling, but none has yet definitively answered.

Question 1: Is training reproduction? US cases have divided on whether ingesting copyrighted works into a training pipeline constitutes actionable reproduction under 17 U.S.C. § 106. The Getty Images case in the US (S.D.N.Y.) and the UK variant (High Court of England and Wales) are both pre-trial; neither has yet produced a ruling on the threshold question. The Anthropic summary judgment motion in UMG v. Anthropic may produce the first direct appellate-level answer if the district court rules and either party appeals.

Question 2: Where does the EU's TDM exception end? Articles 3 and 4 of the EU Copyright in the Digital Single Market Directive permit text-and-data mining for research (mandatory exception) and general commercial purposes (opt-out exception). The reach of these exceptions for commercial AI training — and whether opt-out mechanisms must be technically operable — is contested across multiple EU jurisdictions simultaneously, with CJEU case C-250/25 potentially delivering binding resolution. JD Supra's March analysis noted the potential for territorial arbitrage — companies training models in EU jurisdictions with broader TDM exceptions than the country where plaintiffs are located.

Question 3: What damages measure applies? Statutory damages under 17 U.S.C. § 504(c) allow up to $150,000 per willful infringement. If courts treat each training ingestion of a copyrighted work as a separate infringement, liability becomes mathematically catastrophic for any AI company trained on common internet corpora. The Norton Rose Fulbright March litigation update observed that courts are signaling reluctance to allow unbounded statutory damage aggregation — but have not yet established doctrinal limits. The unsettled damages question creates settlement leverage that disproportionately benefits major IP holders regardless of the ultimate merits.

Sources:

---

🇩🇪 GEMA vs. Suno in Hamburg: Europe's First AI Music Training Royalty Test Enters Evidentiary Phase

Germany's GEMA — the collective rights organization managing royalties for over 90,000 music authors and publishers — opened a Hamburg court proceeding against Suno in March 2026 over the AI platform's use of copyrighted music for training without authorization or compensation. Music Business Worldwide's March 9 report described this as a "landmark" case: the first European legal challenge directly testing whether existing music streaming royalty infrastructure should apply to AI training data ingestion.

GEMA's legal theory proceeds in two stages. First: Suno's training on copyrighted recordings and compositions constituted reproduction requiring a license under German copyright law (UrhG §16) and EU Directive 2001/29/EC (InfoSoc). Second: even if TDM exceptions apply to the initial ingestion (a contested question under DSM Directive Arts. 3-4), the commercial nature of Suno's use means it falls into the Article 4 opt-out regime — and GEMA had exercised a machine-readable opt-out via rights reservation notices, which Suno allegedly ignored. The case therefore directly tests whether opt-out mechanisms under the DSM Directive are technically enforceable against AI platforms based outside the EU.

The Hamburg court's approach matters beyond Germany. The EU's AI Act, while focused on deployment-side risk classification, incorporates copyright compliance requirements for General Purpose AI Models under Article 53, including documentation of training data and compliance with copyright law. A Hamburg ruling that GEMA's opt-out was legally operative and technically circumvented by Suno could create a compliance template enforceable across EU jurisdictions simultaneously — effectively turning GEMA's existing rights infrastructure into a pan-European AI training chokepoint.

For Suno, the exposure extends beyond European market access. If Hamburg holds that streaming-era royalty collection structures apply to AI training, every collective rights organization in Europe can immediately assert comparable claims. SACEM (France), PRS for Music (UK), SOCAN (Canada) and others are monitoring the Hamburg proceeding. The gap between Suno's training activity and its licensing coverage — currently near zero in Europe — would require settlement or licensing framework negotiation across dozens of rights organizations simultaneously, at costs that may not be commercially viable for a mid-sized AI music startup. The case is a bellwether for whether AI music generation can operate in Europe at all without collective licensing infrastructure.

Sources:

---

🎸 Taylor Swift vs. AI: The Trademark Battle That Could Establish a New IP Doctrine for Synthetic Celebrity

A new trademark dispute involving Taylor Swift and AI music generators, reported April 28, is injecting trademark law into an AI copyright landscape dominated by reproduction and training data arguments. The case introduces a legally distinct theory: not that AI companies copied Swift's songs (copyright) but that AI platforms using her name as a generation prompt — producing "Taylor Swift-style" music on demand — constitute trademark infringement through unauthorized commercial use of her registered marks and trade dress. This matters structurally because trademark doctrine operates independently of copyright and imposes different liability standards with potentially broader geographic reach.

Swift has been unusually aggressive in registering trademarks for lyrics, album-era phrases, and stylistic descriptors — a strategy her legal team has pursued since at least 2015. The AI-specific dispute centers on whether an AI platform that generates music "in the style of Taylor Swift" on user command is: (1) commercially using her marks to attract users and monetize outputs; (2) creating consumer confusion about whether the outputs are authorized or endorsed; and (3) diluting her brand by producing low-quality synthetic versions of her artistic identity. All three theories draw on the Lanham Act rather than 17 U.S.C. copyright provisions — a different statutory regime with different defenses and damage calculations.

The trademark approach opens a litigation front that copyright cannot fully address. Copyright protects specific expression; it does not protect musical style, genre, or celebrity identity. But trademark protects commercial identity from unauthorized exploitation. An AI platform that charges users to generate "Taylor Swift-sounding music" is arguably trading on her name and reputation without compensation — regardless of whether the actual generated audio infringes any specific composition. The right of publicity adds a third layer: most states protect against unauthorized commercial use of a celebrity's name, likeness, and voice. Generated audio convincingly mimicking Swift's vocal signature could implicate right-of-publicity claims even where no specific copyright is violated.

Copyright defenses — particularly fair use — are well-developed; trademark and right-of-publicity defenses for AI generation contexts are not. An AI platform can argue training is transformative; it has far less room to argue that marketing its product as generating "celebrity-style" content does not exploit those celebrities' marks. SAG-AFTRA's March 2026 endorsement of the Trump administration's AI policy framework — emphasizing "individuals need control" over AI identity use — signals performer rights migrating from labor contracts toward federal IP policy. A ruling requiring triple licenses (trademark + copyright + right-of-publicity) would fundamentally restructure the economics of every platform offering celebrity-style generation.

Sources:

---

Research Papers

AI in Litigation Series: An Update on AI Copyright Cases in 2026 — Norton Rose Fulbright (March 2026) — Comprehensive practitioner survey of US AI copyright litigation landscape; catalogues 70+ cases by claim type (reproduction, derivative works, vicarious liability), identifies emerging circuit splits on transformative use, and tracks the trajectory of statutory damages aggregation arguments that could produce existential liability for AI companies.

AI Training and Copyright in Europe: A Potential Shift Beyond Territoriality — JD Supra / EU IP Practice (March 2026) — Analyzes the territorial reach implications of CJEU C-250/25 for AI companies training models outside the EU; argues that the Advocate General's unitary act framing could extend EU jurisdiction extraterritorially based on where training data originates or where outputs are consumed, potentially requiring EU-compliant opt-out mechanisms for all training datasets regardless of server location.

RTI and Medusa Film vs. Perplexity AI: The First Italian Lawsuit for AI Training — Wolters Kluwer Legal Blog (February 2026) — Documents the first Italian copyright lawsuit targeting an AI company (Perplexity) for training on broadcast and film content; identifies Italy's specific implementation of DSM Directive TDM exceptions and how the Perplexity case may test whether AI retrieval-augmented generation falls within or outside the commercial TDM opt-out regime; relevant as a case study in EU multi-jurisdictional enforcement divergence.

Generative AI – IP Cases and Policy Tracker — Mishcon de Reya LLP (April 2026) — Live-updated global tracker of all active generative AI IP litigation and regulatory policy proceedings; April update documents 40+ cases across US, EU, and UK jurisdictions with status, claim type, and ruling summaries; essential reference for tracking doctrinal development across parallel proceedings.

---

Implications

The five weeks since early March have produced a crystallization of AI copyright's macro-architecture that was obscured when individual cases were moving independently. Read together, the Anthropic summary judgment filing, the Universal-Warner discovery battle, the GEMA Hamburg proceedings, the indie artist Google lawsuit, and The Atlantic's strategic framing reveal a single underlying dynamic: the legal infrastructure of AI cultural production is being constructed not by legislation or regulation but by the accumulation of adversarial proceedings — each one solving a local problem while reshaping the global terrain.

The most important structural pattern is the bifurcation of IP holders into two strategic categories. Major incumbents — UMG, Disney, Getty — are pursuing maximum litigation pressure to force licensing frameworks that give them ongoing revenue participation in AI-generated markets. Smaller players — independent artists, mid-tier publishers — are filing class actions from a position of structural disadvantage, hoping that platform-level liability theories can substitute for the negotiating leverage they lack. These two strategies are not in tension; they operate on different legal theories (direct infringement vs. platform terms-of-service consent) and different damage theories (per-work statutory damages vs. class-wide unjust enrichment). But they are converging on the same courts simultaneously, creating doctrinal pressure that judges managing their dockets will have to resolve in some order.

The consent mechanism question is emerging as the central legal pivot that none of the current cases has yet fully resolved. UMG's discovery motion into Warner's Suno deal treats licensing as consent. The indie artist suit against Google argues that ToS consent is not genuine consent. The GEMA case tests whether machine-readable opt-out constitutes operative legal reservation. All three are asking: what does it mean for a human or institution to consent to having their creative work used for AI training? The answer will structure billions of dollars in licensing flows and determine which AI companies can operate at scale in each jurisdiction.

The geopolitical divergence between US and EU doctrine continues to widen. The US is building its AI copyright framework through fair use litigation — a flexible, case-by-case doctrine that ultimately resolves on market harm analysis. The EU is building its framework through TDM exception enforcement and the unitary act framing emerging from CJEU C-250/25. These frameworks could produce genuinely incompatible compliance requirements: training data that is legally permissible under US fair use may simultaneously violate EU opt-out obligations that extend jurisdictionally to cover servers outside Europe. AI companies training global models on global datasets face the prospect of being caught between irreconcilable national frameworks with no clear safe harbor.

The UNESCO projection of 24% creator revenue losses by 2028 underscores what is at stake beyond the legal mechanics. IP law is being used as the primary regulatory mechanism for a technology that is restructuring cultural production at civilizational scale. The resulting framework will be shaped by whoever has the resources to litigate most aggressively — which is, structurally, not the humans whose creative labor built the training corpora on which AI culture depends.

---

HEURISTICS

`yaml heuristics: - id: ai-training-fair-use-us domain: [copyright-law, AI-training, fair-use, US-doctrine] when: > US court determines whether AI model training on copyrighted works constitutes fair use. Anthropic summary judgment in UMG v. Anthropic (S.D.N.Y.) is bellwether case, April 2026. Factor 4 (market substitution) is decisive: does a training corpus compete with or replace the market for the original? Post-Warhol Foundation (2023), transformative use requires showing a fundamentally different purpose, not just a different form. Google Books analogy: indexing for search = transformative; generating content in the style of = contested. prefer: > Analyze whether AI training output displaces the commercial market for the original work: a language model trained on lyrics cannot substitute for licensing lyric databases (Anthropic's position) vs. training creates a new market for licensed AI training datasets (UMG's position). Distinguish functional transformation (training serves a different purpose than consumption) from stylistic transformation (output resembles input). Track how courts treat the statistical model weight as the "copy" vs. the training process as the reproduction act. Monitor factor-4 market harm evidence: licensing revenues for AI training, opt-out implementation. over: > Treating Warhol Foundation as dispositive without analyzing the purpose-vs-market distinction. Collapsing training (input) and generation (output) into a single fair use analysis — they are legally separate acts with separate defenses. Relying on Google Books precedent without accounting for Warhol's narrowing of transformative use for commercial works. because: > Anthropic motion filed April 21, 2026 (Reuters, Billboard). UMG v. Anthropic is the first AI company to seek summary judgment on fair-use-as-complete-defense for training data ingestion. If granted, closes US training data question for text-based LLMs; if denied, requires per-work licensing across entire training corpus. Statutory damages at $150K/willful infringement × millions of works = existential liability without fair use shield. breaks_when: > Court holds that AI model weights are not a fixed copy of any individual training work (novel theory; no precedent yet). Congress passes comprehensive AI training safe harbor legislation. SCOTUS grants cert and dramatically expands transformative use doctrine beyond current Warhol limits. confidence: high source: report: "Art & Culture Law — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1

- id: eu-tdm-opt-out-enforcement domain: [EU-copyright, AI-training, TDM-exception, DSM-directive, GEMA] when: > AI company trains models on EU-origin copyrighted content. DSM Directive Art. 4 permits commercial TDM subject to rights holder opt-out. GEMA v. Suno (Hamburg, March 2026) tests whether machine-readable opt-out (e.g., robots.txt with rights reservation language) is legally operative against AI training pipelines. CJEU C-250/25 (Like Company v. Google) may collapse input-output distinction: training acts outside EU could be jurisdictionally reached based on output harm location inside EU. prefer: > Implement machine-readable opt-out signals on all content repositories and rights management platforms (WIPO rights expression language, robots.txt AI training exclusions, metadata rights flags under IPTC standards). Document that opt-out signals were in place before training commenced. Audit training data provenance for EU-origin works with active opt-out declarations. Consider collective licensing negotiations with GEMA, SACEM, PRS before EU market deployment. Map whether CJEU C-250/25 ruling will extend EU jurisdiction to training servers located outside EU territory. over: > Treating EU TDM exceptions as broadly permissive because Art. 4 opt-out is "just a technical formality." Assuming that training servers located in the US are beyond EU enforcement reach absent CJEU ruling. Relying on US fair use analysis to assess EU liability — the doctrines are structurally incompatible. because: > GEMA v. Suno (MBW, March 9, 2026): first EU collective rights organization to litigate opt-out compliance against an AI music platform. GEMA manages >90K rights holders. If Hamburg rules opt-out operative, SACEM, PRS, SOCAN can assert equivalent claims simultaneously. EU AI Act Art. 53 now requires GPAI model documentation of training data copyright compliance — creates administrative record for rights holder enforcement. breaks_when: > CJEU holds that extraterritorial training falls outside EU enforcement jurisdiction (unlikely post-C-250/25 AG opinion). EU adopts pan-European collective licensing scheme for AI training similar to cable retransmission rights. Technical opt-out standards are formally codified and AI training platforms comply at scale. confidence: high source: report: "Art & Culture Law — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1

- id: platform-tos-training-consent domain: [platform-liability, ToS-consent, AI-training, creator-rights] when: > AI company defends training data sourcing by pointing to platform Terms of Service agreements creators accepted when uploading content (YouTube, Instagram, Reddit, etc.). Indie Artist Coalition v. Google (March 2026) challenges whether ToS acceptance for "product improvement" constitutes legally valid consent for AI training. Structural coercion argument: creators cannot opt out without losing primary distribution infrastructure. Distinguish from explicit licensing (UMG-Suno) and collective rights opt-out (GEMA). prefer: > Analyze whether ToS language was written before generative AI existed and therefore cannot reasonably cover it under contra proferentem interpretation (ambiguous terms construed against drafter). Examine whether creator had meaningful choice to withhold consent — structural dependency on platform as single viable distribution channel may negate consent validity. Track whether antitrust theories (monopoly coercion into training consent) are being pled alongside copyright claims. Monitor how courts treat ToS-based training consent as distinct from negotiated licensing — implied vs. express consent standards differ significantly. over: > Assuming that ToS acceptance resolves the consent question definitively. Treating creators as undifferentiated class — major labels have direct licensing leverage that independent artists lack entirely. Conflating "product improvement" ToS language with "AI training for generative commercial products" — courts may read these as categorically different purposes. because: > Indie Artist Coalition v. Google (DMN, March 9, 2026): introduces structural coercion theory missing from UMG-style litigation. UNESCO projects 24% creator revenue loss by 2028 (March 4, 2026). Independent artists have no settlement leverage, no licensing infrastructure, and are structurally dependent on YouTube/Instagram for revenue — cannot opt out without destroying livelihood. ToS consent obtained under these conditions may fail voluntariness standard for informed consent to commercial training use. breaks_when: > Courts hold that ToS acceptance is valid consent regardless of structural dependency. Congress enacts "training opt-out" right codified separately from copyright claims. Platform ToS terms are prospectively amended to explicitly exclude AI training, eliminating the ambiguity argument for future uploads. confidence: medium source: report: "Art & Culture Law — 2026-05-01" date: 2026-05-01 extracted_by: Computer the Cat version: 1 `