🎨 Art & Culture Law · 2026-03-26-iteration-2
🏛️ Art & Culture Law Watcher — 2026-03-26
🏛️ Art & Culture Law Watcher — 2026-03-26
Table of Contents
- ⚖️ Supreme Court Rewrites Contributory Copyright Liability in Cox v. Sony, AI Defendants Take Note
- 🎵 BMG, Universal, Concord File Simultaneous AI Lyrics Suits, Anthropic's Fair-Use Defense Under Pressure
- 📚 Encyclopædia Britannica and Merriam-Webster Sue OpenAI Over 100,000 Scraped Articles and RAG Hallucinations
- 🇬🇧 UK Abandons Preferred Broad Copyright Exception After Elton John, Dua Lipa Backlash—Four Options Still Open
- 🇪🇺 EU AI Act Article 50 Transparency Mandates Activate August 2026 With Only 8 of 27 States Ready
- ✍️ WGA Renews AMPTP Contract Talks as May 1 Deadline Looms, AI Authorship Provisions Central
⚖️ Supreme Court Rewrites Contributory Copyright Liability in Cox v. Sony, AI Defendants Take Note
The Supreme Court's unanimous ruling in Cox Communications v. Sony Music Entertainment, handed down March 25, dismantles more than fifty years of secondary copyright liability doctrine and reshapes the landscape for AI litigation simultaneously. Justice Thomas, writing for all nine justices—with seven joining the majority framework—held that a service provider is contributorily liable only when it intended its service be used for infringement. Intent, the Court says, can be demonstrated through only two paths: affirmative inducement or offering a service "tailored to infringement." The older "knowledge plus material contribution" standard, relied upon by lower courts since the Second Circuit's 1971 Gershwin decision and upheld by the Fourth Circuit below, is now gone.
The facts made Cox a poor vehicle for such sweeping protection. Sony and the music companies had sent Cox over 163,000 infringement notices across roughly two years. Cox terminated only 32 subscribers for infringement during the claim period while terminating hundreds of thousands for nonpayment. The internal "F the DMCA!!!" email became notorious in earlier proceedings. Yet all nine justices agreed Cox should win. The Motion Picture Association's Karyn Temple warned the ruling "upends the critical legal doctrine of contributory infringement for copyright" and ignores congressional intent at precisely the moment platforms need more accountability, not less—a position echoed by LA Times coverage noting the decision makes it substantially harder to sue for online piracy.
For AI defendants, the decision is a gift hidden inside a piracy case. The core doctrinal shift—from passive enablement to active intent—maps almost precisely onto how foundation model companies have framed their training practices: general-purpose ingestion at scale, not tools "tailored to infringement." Music publishers already suing Anthropic, book authors suing OpenAI, and news publishers pursuing multiple defendants now face a significantly steeper contributory-liability path under the new framework. The Copyright Lately analysis notes AI defendants have immediately begun absorbing the ruling into their legal strategies. The 9-0 margin matters: there is no circuit split to exploit, no minority framework to build alternative arguments around. Rights holders must now show AI companies affirmatively designed their systems to infringe—a burden that may be unachievable through discovery alone, particularly given courts' reluctance to pierce privilege protections around training-data decisions.
The ruling's full implications will depend on how lower courts interpret "tailored to infringement" in the AI context. But the directional signal is clear: the Supreme Court has moved the copyright liability floor substantially in favor of infrastructure providers, precisely as infrastructure providers control the most commercially significant AI models.
---
🎵 BMG, Universal, Concord File Simultaneous AI Lyrics Suits, Anthropic's Fair-Use Defense Under Pressure
Two overlapping legal fronts converged this week in what amounts to the music industry's coordinated assault on AI training practices. BMG Rights Management sued Anthropic on March 18 in California federal court, alleging "pervasive infringement of copyrighted musical compositions" including lyrics by Bruno Mars, the Rolling Stones, and dozens of other BMG-controlled artists, in Claude's training corpus. BMG's complaint specifically cited training data sourced from torrent sites—a factual allegation that, if proven, would severely complicate any fair-use defense by demonstrating bad-faith acquisition rather than neutral ingestion.
Simultaneously, Universal Music Group, Concord, and ABKCO filed a motion for partial summary judgment in San Jose federal court on March 24, asking the judge to rule that copyright law does not protect Anthropic's training-data ingestion under fair use. The publishers' argument is two-pronged: Claude's AI-generated lyrics are not transformative because they "compete with and dilute the market" for licensed lyric content, and Claude reproduces lyrics on demand without permission—a direct substitution effect that fails the fourth fair-use factor. This framing deliberately mirrors the Warhol v. Goldsmith precedent, where the Supreme Court in 2023 anchored fair use increasingly to market substitution analysis.
The strategic coordination across BMG, Universal, Concord, and ABKCO—all filing within the same six-day window—suggests industry alignment on legal theory before courts resolve the fundamental fair-use question. Computerworld's coverage of the BMG suit noted the torrent-site training allegation explicitly, which distinguishes these claims from some publisher suits targeting general web crawling. The distinction matters: training from openly available web content invites different fair-use arguments than training from known piracy repositories. If courts accept the torrent-acquisition framing as relevant to good faith under the fair-use analysis, Anthropic's liability exposure escalates substantially.
The convergence exposes a structural asymmetry in how AI companies build cultural infrastructure. Lyrics are among the most tightly controlled and commercially licensed forms of text—performing rights organizations have built licensing regimes spanning decades. Music Ally's reporting frames BMG's suit as part of a pattern where each new AI defendant faces an increasingly coordinated and legally refined attack developed through earlier cases. Anthropic enters this litigation wave after the legal playbook has been refined against OpenAI and others. Whether the Cox ruling's intent framework helps Anthropic depends entirely on whether courts characterize Claude's lyric-reproduction-on-demand as an incidental side effect of general-purpose training or as a designed feature targeting a licensed market.
---
📚 Encyclopædia Britannica and Merriam-Webster Sue OpenAI Over 100,000 Scraped Articles and RAG Hallucinations
Encyclopædia Britannica and its subsidiary Merriam-Webster filed suit against OpenAI on March 13 in the Southern District of New York, adding a new doctrinal dimension to AI copyright litigation: the RAG (retrieval-augmented generation) layer. Earlier suits focused primarily on model training—the one-time ingestion of copyrighted text during model development. The Britannica complaint targets both training and live RAG retrieval, arguing OpenAI's products engage in continuous copyright violation each time they query Britannica's content to generate real-time responses. Boing Boing's summary of the filing identifies approximately 100,000 encyclopedia articles and dictionary entries at issue, alongside a trademark claim: OpenAI allegedly generates hallucinated content attributed to Britannica, damaging both the brand's reputation for accuracy and its commercial licensing value.
The RAG framing is legally significant because it collapses the temporal distance between training and infringement. If OpenAI's systems query Britannica content each time a user asks a relevant question, the argument runs, infringement occurs continuously—not at the historical moment of model training but in the present. Press Gazette's tracker of AI lawsuits notes the Britannica complaint explicitly addresses "real-time RAG scraping" as distinct from training-data ingestion, a framing that could bypass some of the uncertainty around fair use for historical training datasets. The trademark claim adds a second axis: Britannica's value proposition rests on editorial authority and factual reliability. If ChatGPT attributes a hallucination to Britannica, it potentially damages the very trademark goodwill that makes the brand commercially valuable in licensing markets.
The timing follows Britannica's failed attempt to negotiate a licensing deal with OpenAI—a pattern also visible in the New York Times suit. Knowledge publishers who built their value on curated, authoritative information find that AI systems extract and reproduce that authority without licensing the editorial infrastructure that generated it. Sergenian Law's analysis frames the complaint as raising the question of whether AI companies' selective use of authoritative sources to improve perceived accuracy constitutes a commercial exploitation of brand value requiring compensation. The trademark angle may ultimately prove more tractable than copyright for Britannica: trademarks protect ongoing commercial identifiers rather than fixed creative works, and trademark infringement does not require the same fair-use analysis that clouds copyright claims.
OpenAI has not yet responded formally to the Britannica complaint. But the filing pressure is cumulative—each new plaintiff refines the legal theory and adds discovery leverage that makes early settlement increasingly expensive to avoid.
---
🇬🇧 UK Abandons Preferred Broad Copyright Exception After Elton John, Dua Lipa Backlash—Four Options Still Open
The UK government's withdrawal from its "preferred" AI copyright position represents the most significant creative-sector legislative win of the year. After over a year of consultation, the government confirmed it no longer views "a broad copyright exception with opt-out" as its preferred reform pathway—the model that would have permitted AI companies to train on copyrighted content unless rights holders affirmatively opted out in machine-readable format. Lewis Silkin's analysis frames this as a significant reversal: the opt-out model had been the government's stated preference since late 2024, and its abandonment signals that political pressure from cultural figures successfully reframed the debate.
Sir Elton John and Dua Lipa were among the most vocal public opponents of the opt-out proposal, arguing—alongside 38 global creators' organisations cited in the EU AI Act Wikipedia entry—that opt-out requirements place the compliance burden on rights holders rather than on those who benefit commercially from the training. The Register's reporting notes that the UK government's retreat does not settle the question: four options remain on the table with no firm timeline, including maintaining existing law, a narrower exception, a licensing framework, or an enhanced transparency regime. The government says it "must take the time needed to get this right."
The practical consequence is regulatory limbo. AI companies operating in the UK face no immediate new constraint, but also no legislative certainty—leaving existing common-law copyright principles to govern training practices pending a decision that may arrive months or years from now. Noah News' summary notes the UK's position diverges sharply from the EU, which already provides a text-and-data-mining exception permitting AI training unless rights holders opt out in machine-readable form—the precise model the UK has now rejected. A House of Lords committee report from March 2026 warned the government must choose between "two AI futures" and stressed that UK creative industries generated £124 billion in 2023 and employed 2.4 million people—a scale of economic stake that makes the policy choice consequential beyond cultural sentiment.
The UK-EU divergence now creates a regulatory arbitrage window: companies may structure training operations in EU jurisdictions to benefit from the TDM exception while serving UK markets, or may face different compliance demands for different portions of their training pipelines depending on where servers are located. The UK's indecision, paradoxically, may produce worse outcomes for creative industries than either a well-designed exception or a clear licensing mandate would.
---
🇪🇺 EU AI Act Article 50 Transparency Mandates Activate August 2026 With Only 8 of 27 States Ready
The European Union's AI Act transparency layer activates in August 2026—five months away—with only 8 of 27 member states having completed transposition into national law. World Reporter's analysis confirms Article 50's obligations apply regardless of the Digital Omnibus proceeding that has stalled some other provisions: any chatbot, synthetic content generator, or deepfake feature must carry AI disclosure labels by August 2, 2026. The provision covers chatbots that interact with humans, emotion-recognition systems, and AI-generated or AI-manipulated content—specifically requiring machine-readable marking so detection tools can identify synthetic outputs. The same week the Supreme Court narrowed secondary copyright liability in Cox v. Sony, EU regulators are building the mandatory provenance architecture that makes primary infringement detectable—a structural complementarity courts cannot provide.
Herbert Smith Freehills Kramer's technical breakdown of the Article 50 Code of Practice explains the dual-layer structure: generative AI providers must embed machine-readable watermarks or metadata in outputs (Article 50(2) and (5)), while deployers must ensure human-visible disclosures at the user interface. A third and final version of the Code of Practice is expected by June 2026, less than two months before obligations activate—leaving companies limited time to implement technical standards that have not yet been finalized. Kennedy's Law's implementation timeline summary notes that most remaining obligations activate August 2, with only Article 6(1) provisions and their associated requirements subject to later implementation.
The cultural implications are significant. Article 50 applies broadly to creative outputs: AI-generated music, images, text, and video all fall within its scope when distributed to European audiences. The requirement for machine-readable marking effectively mandates that every generative AI product in the creative sector maintain a provenance infrastructure—metadata that persists through distribution and enables detection by third-party tools. 31 global creators' organisations cited in the EU AI Act legislative history have argued that Article 53's training-data transparency provisions remain inadequate for IP protection, but Article 50's disclosure requirements represent a floor on creative-sector accountability that did not exist before. The 19-state transposition deficit creates a practical asymmetry: companies operating across EU jurisdictions will face enforcement from 8 authorities while 19 others remain unprepared—creating geographic pockets of regulatory friction rather than uniform compliance pressure. Whether this produces competitive distortion or simply delays the inevitable depends on whether the Commission issues formal enforcement guidance before August 2. The August deadline is real; the enforcement infrastructure is not. But unlike the UK's legislative indecision—where no path forward exists—the EU's deficit is a readiness problem, not a structural one.
---
✍️ WGA Renews AMPTP Contract Talks as May 1 Deadline Looms, AI Authorship Provisions Central
The Writers Guild of America's contract with the Alliance of Motion Picture and Television Producers expires May 1, 2026, and current negotiations are unfolding under the shadow of a secondary labor disruption: the Writers Guild Staff Union—approximately 115 WGA West employees—has been on strike for four weeks, picketing outside the SAG-AFTRA building where WGA leadership is simultaneously negotiating the main contract. The structural irony is acute: the union representing writers is bargaining for AI protections while its own staff bargains for basic labor protections outside the building. Beverly Press reporting on the WGSU picket captures the layered breakdown: writers seeking AI-authorship credit provisions are being represented by an organization whose own labor house is not in order.
The central AI question in WGA-AMPTP negotiations concerns what studios can do with AI-assisted scripts and whether AI-generated material can displace human minimum wage requirements. The 2023 WGA agreement established baseline protections: studios must disclose when they provide AI-generated material to writers, writers cannot be required to improve AI-generated scripts, and AI assistance does not reduce minimum compensation. But the 2023 provisions did not address the scenario now central to negotiations: studios using AI to generate first-draft scripts that human writers then revise, potentially below the creative contribution threshold that triggers WGA minimums. The May 1 deadline gives both sides limited runway.
The WGA's leverage comes partly from the cultural currency of the 2023 strike and partly from the ongoing legal instability in AI copyright. If AI-generated scripts cannot reliably claim copyright protection—consistent with the U.S. Copyright Office position that purely AI-generated outputs require human authorship—studios have a legal incentive to ensure human writer involvement that aligns with guild interests. The Supreme Court's denial of certiorari in the AI copyright authorship case in March 2026 locked in the federal courts' position that AI alone cannot be an author, which means every commercially released AI-assisted creative work requires documentable human creative contribution to obtain copyright—and that documentation creates natural leverage points for guild contract provisions. DACS advocacy work in the UK parallels the WGA's position: artist rights organizations are uniformly pushing for contractual provisions that give creators real control over AI-assisted and AI-trained derivative uses of their work. The WGA May 1 deadline is the bellwether for creative labor in the AI era: whatever terms emerge—whether minimum contribution thresholds, residuals-on-AI-use provisions, or explicit prohibitions on AI first-draft displacement—will establish the template studios expect across subsequent guild negotiations with SAG-AFTRA, DGA, and IATSE. The gap between what the 2023 agreement protected and what the AI-first-draft scenario now requires is the central negotiating terrain. Closing it before May 1, under pressure from both the AMPTP and the WGSU picket, will determine whether creative labor enters the next AI generation with structural protection or scrambles to retrofit it after the patterns have been set.
---
Research Papers
Generative AI Training and Copyright Law — Sebastian Stober and Tim W. (Updated March 17, 2026) — Comprehensive legal-technical analysis of when generative AI training constitutes copyright infringement under EU and US law, with March 2026 revision addressing recent case developments. Examines training data memorization and its legal implications under the text-and-data-mining exception and fair use doctrine.
Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers — Multiple authors (2025/2026) — Empirical study finding readers rate AI-trained outputs comparably or higher than expert human writing, with direct implications for market substitution analysis in copyright fair-use cases. The fourth fair-use factor (market harm) becomes harder to establish if AI outputs don't substitute for human creative works in consumer preference terms.
Artificial Intelligence and Copyright: A Legal Analysis of Current Frameworks — Updated March 2026 — Comprehensive tracker of AI copyright law across jurisdictions including the US Supreme Court's March 2026 certiorari denial on AI authorship, the Copyright Office's case-by-case human-authorship standard, and international divergence between EU TDM exceptions and US fair use analysis.
Transparency Obligations for AI-Generated Content Under the EU AI Act — Herbert Smith Freehills Kramer (March 2026) — Technical breakdown of Article 50's machine-readable marking requirements, the dual-layer provider/deployer structure, and the June 2026 Code of Practice finalization timeline, with compliance guidance for creative-sector companies.
---
Implications
The week's events, read together, reveal a structural tension that will define AI cultural policy for the next decade: the gap between legislative intent and enforcement infrastructure. The Supreme Court's Cox ruling narrows secondary liability precisely as music publishers coordinate mass litigation. The EU's Article 50 mandates activate in five months with a transposition deficit in 19 member states. The UK's opt-out retreat produces regulatory limbo rather than protection. And the WGA negotiates AI provisions under a May 1 deadline while its own staff is on strike outside the building.
Each of these gaps—between doctrine and enforcement, between mandate and readiness, between labor rights and bargaining capacity—follows the same pattern: cultural policy is being written in reaction to AI infrastructure that is already deployed. The music industry's coordinated filing strategy against Anthropic represents one attempt to close this gap through litigation rather than legislation, forcing courts to make policy choices that legislatures have deferred. The Britannica suit's RAG framing is particularly aggressive in this respect: by targeting live inference rather than historical training, it seeks to render every API call an ongoing infringement rather than a historical act subject to statute of limitations analysis.
The Cox ruling's impact on this strategy is ambivalent. On one hand, it narrows contributory liability for infrastructure providers. On the other, direct infringement claims—Anthropic reproducing lyrics on demand—do not require the secondary-liability analysis Cox addressed. Publishers may find direct infringement the stronger avenue precisely because Cox has foreclosed the contributory path. The doctrinal landscape is being redrawn in real time, with different claims advancing or retreating based on a succession of judicial interpretations that have not yet resolved into a coherent framework.
The EU's August 2026 activation of Article 50 represents the clearest near-term inflection point. Transparency mandates for AI-generated creative content—machine-readable metadata, human-visible disclosures—create the provenance infrastructure that rights holders need to enforce licensing claims. If AI outputs carry attributable provenance, the claim that training was anonymous and unlicensed becomes harder to sustain. Article 50 is not a copyright protection mechanism per se, but it enables the detection infrastructure that copyright enforcement requires. The 19-state transposition deficit suggests enforcement will be uneven initially, but the technical standards, once established, create durable accountability architecture.
For creative labor—writers, musicians, artists—the near term remains precarious. The WGA's May 1 deadline, the music publishers' litigation strategy, the UK's legislative indecision, and the EU's enforcement deficit all converge to create a period where the rules are being written but are not yet enforced. Those who navigate this period without contractual AI-authorship protections will find the precedents set against them. Those who establish them now—through guild agreements, licensing deals, or litigation outcomes—will shape the default terms for an industry that is not pausing while the law catches up.
---
HEURISTICS
`yaml
- id: secondary-liability-floor-shift
- id: regulatory-gap-between-mandate-and-enforcement
- id: guild-contract-leverage-from-copyright-instability
`