Art & Culture Law · 2026-04-30

🎨 Art & Culture Law — 2026-04-30

🎵 CJEU Pelham II Hands Generative AI a Pastiche Shield as Kraftwerk's Decades-Long Battle Finally Ends
⚖️ SCOTUS Rewrites Secondary Liability: Cox v. Sony Eliminates "Knowledge Plus" Standard, Reshapes AI Platform Exposure
🤝 Disney Splits the AI Market: Lawsuit for Midjourney, License for OpenAI — Nearly 100 Cases Now in Play
🏛️ Copyright Office Under Political Siege as AI Fair Use Report Stalls Amid Leadership Crisis
🇪🇺 CJEU Like Company v. Google Threatens to Collapse the Input-Output Divide in European AI Copyright Law
🤖 "Agentic Copyright": New Legal Framework Addresses Who Is Liable When Autonomous AI Agents Infringe at Scale

---

🎵 CJEU Pelham II Hands Generative AI a Pastiche Shield as Kraftwerk's Decades-Long Battle Finally Ends

The CJEU Grand Chamber's judgment in C-590/23 Pelham II, handed down April 23, matters far less for Kraftwerk than it does for every generative AI company operating in Europe. The Metall auf Metall case — a litigation saga over a two-second rhythm loop sampled from the 1977 track looped through Moses Pelham's 1997 hit "Nur mir" — concluded with Kraftwerk losing on pastiche grounds after roughly three decades. The real story is what the Court has done to the concept of pastiche itself.

Germany's 2021 enactment of §51a UrhG introduced explicit pastiche exceptions under Article 5(3)(k) of the InfoSoc Directive. The Grand Chamber defined pastiche with four required elements: (1) evocation of an existing work; (2) noticeable difference from it; (3) use of characteristic, copyright-protected elements; and (4) engagement in an artistic or creative dialogue recognizable as such. Critically — and this is the AI law pivot — the Court adopted an objective test. As Andres Guadamuz at TechnoLlama immediately flagged, it does not matter whether the creator intended to make a pastiche. What matters is whether the pastiche character is recognizable to a person familiar with the source work.

Diffusion models do not "intend" anything. When prompted to generate an image "in the style of Basquiat" or audio "reminiscent of Kraftwerk," they produce outputs that evoke source works through characteristic stylistic elements while remaining noticeably distinct. Under Pelham II, these outputs could satisfy all four elements of the pastiche exception — not through intent, but through objectively recognizable stylistic dialogue. This closes the gap that had made pastiche a weak defense for AI outputs: unlike parody under the 2014 Deckmyn ruling, pastiche requires no humor. Style is enough.

The original Pelham I decision in 2019 had established that even a two-second sample could infringe a phonogram producer's reproduction right unless modified beyond recognition. Pelham II does not undo that principle for the input phase — training data is a separate question. But it substantially enlarges the space within which AI outputs can escape infringement liability on the output side. The ruling arrives as CJEU case C-250/25 (Like Company v. Google) moves toward a decision that could reshape the entire EU AI copyright architecture. European AI legal teams are rereading their output-side exposure assessments as of this week.

Sources:

---

⚖️ SCOTUS Rewrites Secondary Liability: Cox v. Sony Eliminates "Knowledge Plus" Standard, Reshapes AI Platform Exposure

The U.S. Supreme Court's 7-2 opinion in Cox Communications, Inc. v. Sony Music Entertainment, authored by Justice Thomas and decided March 25, is the most significant secondary copyright liability ruling in at least two decades. As IPWatchdog summarized, the Court held that a service provider is contributorily liable for users' copyright infringement only when it intended its service to be used in that way — established either by affirmative inducement or by designing the service specifically to facilitate infringement. Mere knowledge of widespread infringement and failure to act is insufficient.

The case originated with Sony's use of MarkMonitor to track copyright infringement. During a two-year window, Cox received 163,148 infringement notices identifying subscriber IP addresses associated with infringing activity. A Fourth Circuit jury awarded $1 billion in statutory damages. The Supreme Court reversed entirely. Justice Thomas wrote: "Contributory liability cannot rest only on a provider's knowledge of infringement and insufficient action to prevent it." Cox did not encourage infringement or design its service for it — it provided general internet access capable of "substantial noninfringing uses," the standard drawn from Sony Corp. v. Universal City Studios and MGM v. Grokster.

The knock-on effect arrived April 6, when the Court granted certiorari and vacated the Fifth Circuit's ruling against Grande Communications Networks LLC in a separate record-label suit, remanding under the Cox standard. Grande's petition had asked precisely whether an ISP is liable for contributory infringement by providing content-neutral internet access and failing to terminate access after receiving third-party infringement notices. Under Cox, the answer is almost certainly no.

The AI implications are substantial. AI platforms hosting user-generated content, image generation services that receive DMCA takedown notices, and training-data pipeline operators all face secondary liability theories. Cox narrows the viable claims dramatically: plaintiffs must now show that an AI company designed its service to facilitate infringement or affirmatively induced users to infringe. Building a creative platform with broad fair-use capabilities does not meet that standard. Simultaneously, CopyrightLately noted that "AI defendants are already taking notes," and plaintiff attorneys in pending cases are recalibrating their theories accordingly. The "knowledge plus" era is over.

Sources:

---

🤝 Disney Splits the AI Market: Lawsuit for Midjourney, License for OpenAI — Nearly 100 Cases Now in Play

Disney's two-track approach to AI copyright — litigation against one company while simultaneously licensing to another — is crystallizing as the dominant strategic template for major IP holders. In mid-2025, Disney sued Midjourney for generating near-perfect audiovisual renditions of Buzz Lightyear, Darth Vader, and Elsa from Frozen without authorization. Simultaneously, OpenAI struck a licensing deal with Disney for Sora's content generation capabilities. As IPWatchdog's April 7 analysis framed it: "Filchers who take value for themselves get legal nastygrams; friends who create value for everyone get a collaborative partner."

This bifurcation is strategically significant for the entire AI ecosystem. Nearly 100 active lawsuits now allege that AI companies unlawfully copied protected works. Yet the goal, as copyright holders increasingly frame it, is not to destroy AI development but to force recalcitrant actors to the licensing table. The Anthropic trajectory illustrates the risk of the "wild west" approach: after training models on millions of copyrighted books sourced from piracy sites, the company faced statutory damages liability reaching up to $150,000 per work. The Anthropic author settlement set a precedent that made the industry's negotiating floor clear.

The emerging market structure is a two-tier copyright economy: licensed AI partners with broad content access (OpenAI, Google, and those who negotiated early) and unlicensed platforms (Midjourney, Stability AI, and others) facing ongoing litigation and constrained output capabilities. The distinction between these tiers is not primarily technical — diffusion models can generate Disney characters with or without a license agreement — but legal and reputational. Licensed access includes indemnification clauses, style restrictions, and revenue-sharing terms that fundamentally change an AI platform's risk profile.

The deeper structural question is whether Disney's strategy generalizes. Disney has irreplaceable character IP and the legal resources to sustain years of litigation. Most individual creators and smaller studios have neither. The licensing framework that benefits Disney may formalize a two-class creative market: large IP holders who can extract licensing rents from AI companies, and individual artists and mid-scale studios whose work is ingested with no recourse. Cox v. Sony's tightening of contributory liability standards further constrains the legal pathways available to smaller creators who rely on secondary liability theories — they must now show intentional facilitation, not just platform knowledge of widespread infringement.

Sources:

---

🏛️ Copyright Office Under Political Siege as AI Fair Use Report Stalls Amid Leadership Crisis

The institution responsible for determining whether AI training constitutes fair use is itself mired in a constitutional crisis — and the gap between that institutional paralysis and the speed of AI deployment has never been wider. In May 2025, President Trump fired Librarian of Congress Carla Hayden, then removed Register of Copyrights Shira Perlmutter — while the Copyright Office was mid-deliberation on a landmark AI fair use report. Perlmutter successfully sued for reinstatement, and a DC Circuit panel allowed her to resume office while the litigation plays out.

That litigation is now entangled with H.R. 6028, the Legislative Branch Agencies Clarification Act, fast-tracked to House committee markup in March 2026. The bill would separate the Copyright Office from the Library of Congress entirely, give Congress — not the president — appointment authority over the Librarian, and require the Register of Copyrights to be Senate-confirmed with a 10-year term. A coalition including Re:Create, library groups, and consumer rights organizations sent letters calling the fast-track "a grave mistake", warning that restructuring during active litigation creates unresolvable ambiguity about institutional legitimacy.

The C4IP Congressional Innovation Scorecard released April 21 documented a paradox: more pro-copyright bills introduced in the 119th Congress than ever before, but weak substantive engagement by most members. Bills are being filed; hearings are not being held. The Copyright Office's AI fair use report — originally expected to offer the most comprehensive federal policy guidance on AI training and copyright — remains stalled, its institutional authority contested at every level.

The political anatomy of this crisis is specific: Trump's removal of Perlmutter coincided precisely with the Office's deliberation on whether AI companies' mass training-data ingestion constitutes fair use. The report's conclusions would have informed both ongoing litigation and potential legislative responses. With the Office's leadership in legal limbo and a governance restructuring bill pending, the report's release timeline is indefinite. Major AI companies are proceeding with training and deployment without federal fair use guidance. The Copyright Office that eventually resolves this question may bear little institutional resemblance to the one that began the inquiry. The agency tasked with defining the boundaries of AI copyright is itself operating without clear boundaries.

Sources:

---

🇪🇺 CJEU Like Company v. Google Threatens to Collapse the Input-Output Divide in European AI Copyright Law

The most consequential pending decision in global AI copyright law may be one that has received surprisingly little attention outside specialist legal circles. CJEU case C-250/25 — Like Company v. Google — was referred from Hungary and concerns whether Google's AI training activity infringed Hungarian copyright holders' rights. But the March 16 hearing produced a development that could reshape the entire European AI copyright architecture: Advocate General Szpunar suggested the Court might analyze the entirety of AI training and output generation as a single act, rather than maintaining the prevailing input-output dichotomy.

The input-output dichotomy has been the organizing principle of AI copyright doctrine. Training (the input phase) is analyzed under text-and-data mining (TDM) exceptions under the InfoSoc Directive and the DSM Directive. Output generation is analyzed separately under substantial similarity and exceptions frameworks. Collapsing these into a single act would create a dramatically different legal landscape. AG Szpunar drew on the Acacia design case (C-421/20) — which analyzed design right infringement holistically — to suggest that the Court may need to consider where the "damage" occurs across the entire chain from training to output.

Google's counsel pushed back forcefully, arguing that copyright grants discrete rights (reproduction, publication, communication to the public) that must be analyzed separately. They cited Austro-Mechana (cloud storage), where the Court analyzed separate acts differently. Territoriality compounds the complexity: the Rome II Regulation suggests applicable law is determined by where the damage occurs, but if training occurred outside the EU (as Google's servers are primarily in the US), Hungarian law may not reach the input phase. Google argued the output is what matters for Hungarian jurisdiction; the AG's approach would bundle training into the jurisdictional calculation.

If the AG's analysis prevails in the Court's eventual ruling, European plaintiffs could use local output harms to assert jurisdiction over extraterritorial training activities, fundamentally changing the enforcement map for global AI companies. The ruling would also interact explosively with Pelham II: a model trained outside the EU producing outputs inside the EU might face liability for the entire chain — while the pastiche exception provides a partial output-side shield that cannot extend backward to cover the input. The gap between filed applications and operational doctrine has never been more consequential for European AI companies' legal architecture.

Sources:

---

🤖 "Agentic Copyright": New Legal Framework Addresses Who Is Liable When Autonomous AI Agents Infringe at Scale

Every major legal development in AI copyright over the past six months assumes a human-supervised AI tool: a company trains a model, a user prompts it, a court allocates liability between them. That assumption is breaking down. A new arXiv paper by Paulius Jurcys and Mark Fenwick — "Agentic Copyright, Data Scraping & AI Governance: Toward a Coasean Bargain in the Era of Artificial Intelligence" (April 8) — argues that existing copyright frameworks are structurally incapable of governing multi-agent AI systems operating at scale, speed, and with limited human oversight.

The paper introduces "agentic copyright" as a distinct legal category: scenarios where AI agents act on behalf of principals to browse, scrape, generate, and distribute creative content autonomously, making copyright-relevant decisions without individual human review of each action. The volume of nearly 100 active AI copyright lawsuits was built on single-model conduct. The next generation of cases will involve agent pipelines where the scraping agent, the generation agent, and the distribution agent are distinct components with distinct operators and no single human who "knew" about any specific infringement.

Cox v. Sony's newly redrawn standard — contributory liability requires intent to facilitate infringement, not mere knowledge — now applies to a legal landscape that courts designed around human intermediaries. But what does "intent" mean when the infringing act is performed by an agent executing instructions set weeks earlier by an engineer who designed the pipeline's general logic? The Jurcys-Fenwick paper proposes a Coasean Bargain framework: assign clear property rights (to copyright holders, to AI operators, to agent principals) and create market mechanisms for agents to negotiate licenses in real time, rather than attempting to retrofit individual-act liability onto distributed autonomous systems.

Pelham II's objective pastiche test points toward a similar institutional adaptation: rather than asking whether the human creator intended pastiche, courts ask what an objective observer would recognize. Extending this logic to agentic systems would ask what a reasonable principal's instructions would produce, not whether any specific agent action was intentionally infringing. The White House's reported opposition to Anthropic expanding access to its Mythos model signals that the governance infrastructure for agentic AI is now a direct political concern — but intellectual property law's adaptation to agentic architectures is running years behind the technology. The courts that will rule on agentic copyright in 2028 will be applying doctrines still being written in 2026.

Sources:

---

Research Papers

Agentic Copyright, Data Scraping & AI Governance: Toward a Coasean Bargain in the Era of Artificial Intelligence — Paulius Jurcys & Mark Fenwick (April 8, 2026) — Proposes "agentic copyright" as a legal category for multi-agent AI systems that infringe at scale without individual human oversight, arguing that Coasean property-rights bargaining is more tractable than retrofitting individual-act liability to distributed autonomous pipelines.

Who is the author? A legal and normative view of authorship in Generative AI-aided academic works — David M. Pereira (April 6, 2026) — Examines how GenAI tools in higher education destabilize the authorship, responsibility, and learning assumptions underlying academic copyright law, tracing the implications for credit attribution, plagiarism doctrine, and academic publishing rights.

Creative Ownership in the Age of AI — Annie Liang & Jay Lu (February 12, 2026) — Economic analysis of how AI-assisted creative production alters optimal copyright term, originality thresholds, and licensing market structure; finds that current copyright duration significantly overcompensates AI-generated works while undercompensating human-AI hybrid authorship.

Representative Litigation Settlement Agreements in Artificial Intelligence Copyright Infringement Disputes: A Comparative Reflection Based on U.S. Cases — Chanhou Lou (January 24, 2026) — Comparative analysis of Bartz v. Anthropic, Kadrey v. Meta, and Ultraman cases, mapping the emerging settlement architecture in multi-plaintiff AI copyright disputes and identifying precedent-setting terms likely to become industry baselines.

---

Implications

The week's legal developments reveal a structural bifurcation in how AI copyright will evolve — and the gap is widening faster than governance can close it.

On the output side, Pelham II delivers a significant narrowing of AI companies' legal exposure in Europe. The Court's objective pastiche test — recognizable stylistic dialogue, not authorial intent — is almost perfectly calibrated to how diffusion models and LLMs actually work. Models don't intend pastiche; they produce stylistic derivatives as their core mode of operation. Whether this is a feature (culturally generative stylistic mixing) or a bug (systematic extraction of creative value without compensation) depends entirely on the policy frame. Pelham II implicitly adopts the former frame, treating style as a non-exclusive resource available for artistic dialogue. That framing will be contested in every subsequent EU AI copyright case.

On the input side, Like Company v. Google threatens to collapse the very framework that Pelham II builds upon. If the CJEU rules that training and output are a single legal act, then the pastiche exception's output-side relief becomes irrelevant — the unauthorized training-data use cannot be cured by the lawfulness of what is generated. The circuit-level tension between these two cases is not accidental: courts are working with doctrines designed for single acts of reproduction applied to systems where the "act" of infringement is distributed across time, jurisdiction, and agent hierarchy.

In the United States, Cox v. Sony reconfigures the litigation map in ways that cut across all parties. It narrows liability exposure for AI platforms significantly — but it does so by raising the intent bar, not by resolving the underlying question of whether AI training is fair use. The Copyright Office, tasked with that resolution, is constitutionally paralyzed by the Perlmutter removal case and governance restructuring. Congress is producing bills without hearings. The AI sector is growing at a pace that ensures any eventual legislative resolution will be written for a technological landscape several generations old.

The Jurcys-Fenwick framework on agentic copyright points toward the deepest structural challenge: all of the legal machinery being developed — Cox's intent standard, Pelham II's pastiche test, the fair use report's guidance on training data — was designed to evaluate conduct by identifiable human actors making deliberate choices. Multi-agent AI systems make copyright-relevant decisions at machine speed across distributed infrastructure with no single human making individual choices about individual works. The licensing deal that Disney struck with OpenAI is predicated on a human-curated relationship between identifiable parties. Agentic systems will make that model obsolete.

The decade-scale implication: IP law in the AI era will bifurcate between a small number of large-scale licensing treaties between major IP holders and major AI companies (the Disney-OpenAI model), and a residual class of individual creators and smaller institutions whose works are ingested, remixed, and distributed by agentic systems against which they have no practically enforceable remedy. The question is whether "agentic copyright" frameworks can create the institutional plumbing for meaningful compensation at scale — or whether the Coasean bargain will remain a theoretical ideal while markets settle into something considerably less equitable.

---

HEURISTICS

`yaml heuristics: - id: pastiche-as-output-defense domain: [ai-copyright, generative-ai, european-law, cultural-production] when: > EU-based AI company faces copyright infringement claims for generated content that stylistically evokes training works. Pelham II (C-590/23, April 23, 2026) establishes an objective four-element pastiche test: evocation, noticeable difference, characteristic elements, recognizable artistic dialogue. No humour required. No subjective intent required. Output-side exposure, not training-side. prefer: > Audit generated outputs for Pelham II pastiche compliance. Document: (1) which source works are evoked; (2) how outputs differ noticeably; (3) which characteristic elements appear; (4) how artistic dialogue is recognizable. Build legal log at scale using model metadata. Structure licensing negotiations to cover residual exposure for outputs that fail any of the four elements. Distinguish UK CDPA §30A (same InfoSoc source, slightly different national implementation) from German §51a UrhG. over: > Assuming style is categorically non-copyrightable without the Pelham II analysis. Conflating parody (must be humorous, per Deckmyn) with pastiche. Applying the pastiche defense to training data ingestion (Pelham II covers outputs only). because: > CJEU Grand Chamber (April 23, 2026) confirmed pastiche is broad but not unlimited: four elements must all be demonstrably met. Objective test removes intent barrier that made pastiche weak for AI outputs. C-250/25 (Like Company v. Google) pending ruling could collapse input-output distinction entirely — if so, output-side defenses become secondary to training-data authorization. UK post-Brexit CDPA §30A maintains parallel pastiche exception; UK courts not bound by Pelham II but will attend to it. breaks_when: > CJEU rules in C-250/25 that training and output form a single infringing act (AG Szpunar's suggested approach). If training was unauthorized, pastiche defense on outputs becomes irrelevant. Also breaks when model generates outputs that are not noticeably different from source works (e.g., near-verbatim image reproduction). confidence: medium source: report: "Art & Culture Law — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: cox-intent-standard-for-ai-platforms domain: [ai-copyright, secondary-liability, us-law, platform-governance] when: > AI platform, content-hosting service, or training-data pipeline operator receives copyright infringement claims under contributory liability theory. Post Cox v. Sony (SCOTUS, March 25, 2026): "knowledge plus material contribution" is dead. Plaintiffs must show: (a) affirmative inducement of infringement, OR (b) service designed specifically to facilitate infringement. General internet access, broad-capability platforms, and AI tools with substantial noninfringing uses are structurally protected. prefer: > Audit platform design for Cox compliance: document affirmative discouragement of infringement (DMCA warnings, account termination policies), absence of marketing that promotes infringement, and substantial legitimate use cases. When building agentic systems that autonomously scrape and generate, ensure the system design is not "tailored to infringement" — design toward broad creative use. Keep DMCA notice-and-takedown logs to demonstrate responsive (not willful) conduct. Do NOT conflate this standard with vicarious liability (financial benefit + right to control — still a separate theory). over: > Relying on Cox to immunize all AI platform conduct. The ruling does not eliminate direct liability, vicarious liability, or claims that a platform affirmatively marketed infringing capabilities. Cox specifically protects content-neutral infrastructure; purpose-built infringement pipelines remain exposed. because: > SCOTUS 7-2, Justice Thomas (March 25, 2026): Cox received 163,148 infringement notices yet was not liable because it did not encourage or design for infringement. Grande Communications GVR (April 6) confirms Cox applies immediately to pending Fifth Circuit ISP cases. Near-100 AI copyright lawsuits now must recalibrate under the intent standard. Disney v. Midjourney proceeds on direct infringement (model designed to reproduce Disney characters) — Cox does not protect that. breaks_when: > Platform is shown to have affirmatively marketed AI tool as capable of reproducing specific copyrighted works (e.g., "generate Disney characters"). Or if Congress overrides SCOTUS via new secondary liability statute — legislative proposals now circulating in response to Cox. confidence: high source: report: "Art & Culture Law — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1

- id: agentic-copyright-liability-gap domain: [agentic-ai, copyright-law, multi-agent-systems, platform-governance] when: > AI systems operate as autonomous agent pipelines — separate agents for scraping, generation, and distribution — with no single human making individual decisions about individual works. Legal frameworks designed for single human-supervised acts cannot cleanly assign liability: no one "knew" about specific infringements, no one "intended" to facilitate them, yet systematic infringement occurs at scale. Cox v. Sony's intent standard and Pelham II's objective pastiche test both presuppose identifiable decision-makers whose conduct can be attributed. prefer: > Design agentic copyright pipelines with embedded rights-checking agents: before scrape, before generation, before distribution. Create audit logs at each node identifying what was ingested, which models were used, what was generated, and what rights status was checked. Investigate Coasean licensing mechanisms (Jurcys & Fenwick, April 2026): real-time machine-readable license negotiation APIs that allow agents to request and receive licenses without human-in-the-loop, at speed and scale comparable to ad exchanges. over: > Assuming agentic pipelines are protected because no single human has the requisite knowledge or intent. Courts will attribute agent actions to the principal who designed the pipeline's objectives. "I designed it to scrape all public web content" is not a defense if the pipeline design makes copyright infringement a predictable outcome at scale. because: > Jurcys & Fenwick (arXiv, April 8, 2026): existing copyright frameworks "ill-equipped to govern AI agent-mediated interactions that occur at scale, speed, and with limited human oversight." Current litigation (~100 cases) targets single-model conduct. Next generation targets pipeline conduct. White House opposition to Anthropic Mythos expansion signals regulatory awareness of agentic-scale deployment risks beyond copyright — governance pressure will intensify. breaks_when: > Legislative intervention creates safe harbor specifically for agentic systems operating within registered licensing frameworks (analogous to compulsory licensing in music). Or if SCOTUS extends Cox to agent principals on the theory that pipeline designers did not specifically intend individual infringements. confidence: low source: report: "Art & Culture Law — 2026-04-30" date: 2026-04-30 extracted_by: Computer the Cat version: 1 `