Art & Culture Law · 2026-04-28

🎨 Art & Culture Law Watcher — 2026-04-28

🎤 Taylor Swift Files Sound and Image Trademarks to Block AI Voice and Likeness Clones
⚖️ Anthropic Claims Song Lyric Training Is "Transformative" in Bid to End $3B Music Publisher Lawsuit
🍁 Canada's Heritage Committee Recommends Opt-In AI Consent — A Policy That Could Erase Canadian Culture from Models
📚 AI Copyright FAQ for Writers Signals Publishing Sector Has Entered the Practical Compliance Phase
🔏 Provenance Marking Emerges as Artists' Last Defense Against AI Authenticity Collapse
🌏 Australia's "AI Theft" Coalition and the Global Fracture in Creator Rights Policy

---

🎤 Taylor Swift Files Sound and Image Trademarks to Block AI Voice and Likeness Clones

On April 24, Swift's company TAS Rights Management filed three trademark applications with the U.S. Patent & Trademark Office: two sound marks covering her voice (the phrases "Hey, it's Taylor Swift" and "Hey, it's Taylor") and one visual trademark covering a specific performance image — iridescent bodysuit, pink stage, purple lighting — described with enough particularity to cover AI-generated likenesses that evoke but don't reproduce it.

The filings, spotted by IP attorney Josh Gerben, follow the playbook pioneered by Matthew McConaughey, who secured eight trademarks in 2025 including a sound mark for "Alright, alright, alright!" McConaughey's legal team developed the "trademark yourself" strategy to create federal-court jurisdiction over AI identity theft claims — right-of-publicity laws, the previous instrument of choice, operate only at state level and provide inconsistent remedies across California, New York, and elsewhere.

The theory is technically elegant. If an AI platform generates audio using Swift's registered voice mark, she can pursue trademark infringement in federal court with nationwide scope — stronger injunctive remedies than state right-of-publicity claims allow, and no requirement to prove a specific copyrighted recording was reproduced. Trademark requires only distinctiveness and consumer confusion. The approach bypasses the AI-copyright authorship debate entirely.

Swift's likeness has been misappropriated repeatedly by AI systems — Meta chatbots adopted celebrity personas without authorization, pornographic deepfakes circulated in 2024, and Trump's 2024 campaign shared AI-generated images falsely claiming her endorsement. The trademark route represents a shift from reactive takedowns to proactive identity infrastructure.

The strategy hasn't yet been tested in court against AI, but a December 2025 Disney cease-and-desist to Google over Gemini generating trademarked characters produced removal within 24 hours — the kind of compliance speed that multi-year copyright litigation cannot achieve. Gerben's analysis: if an AI generates voice that "sounds like the registered trademark," Swift can claim trademark infringement. The visual trademark, covering a specific jumpsuit-and-stage configuration, creates grounds to pursue claims against AI-generated images that evoke her likeness without reproducing any copyrighted photograph.

This marks a categorical shift. Copyright governs fixed creative works; trademark governs signifiers of origin. When AI generates a voice that sounds like Taylor Swift, the operative question is no longer "did it copy a recording?" but "does this create confusion about whose voice it is?" — which is trademark's native jurisdiction. As AI voice and image synthesis becomes frictionless, trademark law is being conscripted as primary enforcement infrastructure for human identity in digital environments, covering territory copyright was never designed to reach.

Sources:

---

⚖️ Anthropic Claims Song Lyric Training Is "Transformative" in Bid to End $3B Music Publisher Lawsuit

Anthropic filed for summary judgment this week in a consolidated music publisher lawsuit, arguing that training Claude on copyrighted song lyrics constitutes fair use under Section 107 of the Copyright Act — with "transformativeness" as the centerpiece claim. The case, filed in January 2026 by Universal Music Group and major publishing houses seeking $3 billion over alleged "flagrant piracy" of 20,000 works, has become the most consequential AI training copyright test since the Supreme Court declined in March 2026 to take up AI authorship.

Anthropic's argument — that learning patterns from lyrics (structure, emotional register, cadence) is categorically different from reproduction, and that Claude outputs don't substitute for original works in any market — applies the Campbell v. Acuff-Rose framework: the more transformative the use, the less weight market substitution receives. Billboard headlined the position as "Training on Lyrics Is Transformative," which encapsulates both the legal claim and its commercial stakes for Anthropic's $380B valuation.

The music publishers' counter, reinforced by BMG's March 2026 separate lawsuit alleging infringement of Bruno Mars and Rolling Stones catalog, is that Anthropic's valuation was "built on stolen copyrighted works" — that the training act itself, regardless of output, constitutes direct reproduction. Publishers filed their own summary judgment motion in March, arguing training is not transformative, that market harm is clear (displacing licensed lyric databases and music annotation services), and that in-context lyric reproduction during model operation constitutes separate infringement beyond training.

The case crystallizes the core jurisprudential problem: copyright's architecture assumes human authorship as the entry point and reproduction as the harm. Does training = reproduction? Does the model = a derivative work? Does in-context output = infringement separate from training? US courts have not resolved any of these questions. Warhol Foundation v. Goldsmith (2023 SCOTUS) narrowed transformativeness for commercial derivative works, but AI training is not a derivative work in the conventional sense. A summary judgment ruling against Anthropic could establish that each training exposure to a copyrighted work is a reproduction — potentially collapsing the foundation of every major LLM trained on text.

The dual summary judgment posture, with both sides seeking dismissal on legal grounds, means the court will have to decide the training-reproduction question directly rather than letting it survive to trial. Whichever way the Northern District of California rules, it will move immediately toward appeal. The outcome will determine the liability architecture of the AI industry for the rest of the decade.

Sources:

---

🍁 Canada's Heritage Committee Recommends Opt-In AI Consent — A Policy That Could Erase Canadian Culture from Models

Canada's Standing Committee on Canadian Heritage released its AI and creative industries report this month, and its centerpiece recommendation — an opt-in consent requirement for the use of copyrighted Canadian works in AI training — would position Canada as a global outlier, with consequences that law professor Michael Geist argues would backfire severely.

Under opt-in consent (Recommendation 1(c)), AI developers cannot use any Canadian copyrighted work for training unless the rights holder explicitly consents in advance. Compare this to Article 4 of the EU's Digital Single Market Directive, which permits text and data mining unless rights holders expressly opt out — a pro-training-inclusive default. Japan's Copyright Act permits copyrighted material for data analysis including AI training with no opt-out mechanism. Singapore's Copyright Act creates specific exceptions for computational data analysis including commercial use.

The Canadian proposal's defect is structural. Canadian cultural content represents a small fraction of the global training corpus. If inclusion requires prior written authorization from every rights holder, developers will simply exclude Canadian material. This is what Geist calls the Online News Act paradox: when Canada mandated news licensing fees, Meta blocked Canadian news entirely. The Heritage Committee built its report around 43 witnesses, the overwhelming majority from collective rights organizations — Access Copyright, SOCAN, Music Canada, Music Publishers Canada — and cultural industry associations. The handful of witnesses offering alternative perspectives on competitiveness and AI research were "cited in the report but functionally absent from its recommendations."

The committee adopted the "ART" framework (Authorization, Remuneration, Transparency) promoted by those same collective rights organizations as its organizing principle, while the report's own body text acknowledges the testimony "did not yield a consensus on the exact means of striking the balance between facilitating innovation and protecting Canadian creators." The recommendations resolve every contested question in a single direction, manufacturing consensus from contested terrain.

Canada's copyright law already includes fair dealing provisions that would likely cover most AI training under existing doctrine. The opt-in recommendation doesn't create a new exception — it creates a new barrier that effectively overrides fair dealing by requiring prior authorization not found in the Copyright Act. The irony: a report intended to support Canadian creators may ensure Canadian culture is absent from the AI systems that increasingly mediate how the world encounters information and culture, precisely because it treats the cultural corpus as a licensing object rather than a cultural presence.

Sources:

---

📚 AI Copyright FAQ for Writers Signals Publishing Sector Has Entered the Practical Compliance Phase

When Jane Friedman, whose publishing industry newsletter reaches professional authors and editors worldwide, published an AI and copyright FAQ for writers in late April, it marked a threshold: the legal uncertainty of 2023-2025 has hardened enough into operational guidance that writers can now receive actionable answers rather than theoretical frameworks. The shift from "here's what the law might become" to "here's what you need to do" signals the publishing sector has moved into practical compliance.

The FAQ's core answers are blunt. Using AI to generate text that you submit as your own writing raises both copyright and disclosure issues: the US Copyright Office's AI policy requires disclosure of AI-generated portions and denies copyright protection for fully AI-generated text, while publishers and literary agents have updated contracts to require disclosure of AI assistance and many exclude AI-generated manuscripts entirely. The Authors Guild has issued formal guidance opposing AI-generated content passing as human writing and supporting opt-out mechanisms for training data use.

The practical copyright situation for writers is now clearer than it was two years ago. A work with substantial human authorship that incorporates AI-generated elements can be registered with the Copyright Office with appropriate disclosure — the human-authored portions receive protection, the AI-generated portions do not. The exact threshold of "substantial human authorship" remains undefined, but the Copyright Office's 2024 guidance in the context of Kristina Kashtanova's Zarya of the Dawn (AI-generated images denied protection, human-written text protected) established the core principle.

The publishing sector's policy divergence reveals where the battle has moved. Amazon KDP requires AI-generated content disclosure on self-published works. Wattpad prohibits fully AI-generated stories. Traditional publishers' agent submissions increasingly specify "no AI-generated manuscripts." The WGA's 2023 contract established that studios cannot use AI to generate scripts, only to develop ideas for human writers to "polish" — the first major collective bargaining outcome on AI and authorship, setting a template for labor-side responses that the Authors Guild and American Guild of Authors & Journalists have cited as a model.

The convergence of Copyright Office guidance, collective bargaining outcomes, and publisher contract updates means writers now face a tripartite compliance requirement: disclose AI assistance at the point of copyright registration, comply with union/guild AI restrictions if working in organized sectors, and meet publisher-specific AI disclosure or prohibition terms. The FAQ stage means these requirements have been stable enough to codify. The policy is not converged — publishers differ, guilds differ, platforms differ — but the scaffolding is now visible enough for practical guidance.

The structural implication: the authorship question that seemed entirely theoretical in 2023 ("what happens when AI writes text?") has resolved into an operational question ("how much AI assistance disqualifies your work from copyright protection or publication eligibility?"). The legal ambiguity has not been resolved by courts; it has been resolved by contracts, guild agreements, and administrative guidance, the faster-moving institutional mechanisms that fill the gap between legislative timelines and market practice.

Sources:

---

🔏 Provenance Marking Emerges as Artists' Last Defense Against AI Authenticity Collapse

As AI-generated content saturates distribution channels, artists and institutions are turning to provenance marking — embedding cryptographic metadata into human-created works to distinguish them from synthetic outputs — as a structural defense against attribution collapse. BBC coverage this month highlighted a striking example: the famous "monkey selfie" photograph, whose status as definitively non-human-authored makes it an ideal provenance anchor for authentication systems. A work whose legal non-copyrightability is settled fact needs no digital signature to establish its non-AI origin.

The Naruto v. Slater case established in US courts that animals cannot hold copyright, placing the macaque's 2011 self-portrait irrevocably in the public domain with legally unambiguous provenance — making it a baseline calibration tool for AI detection systems that test whether authentication frameworks can identify content whose origin is definitively known. Researchers are exploiting this legal certainty as a ground truth marker in systems designed to detect AI modification.

The technical infrastructure for human-created content is the Coalition for Content Provenance and Authenticity (C2PA), a joint standard developed by Adobe, Microsoft, Google, Intel, and publishers. C2PA embeds cryptographically signed metadata into media files at creation, recording the tool used, the date, and the creator's identity. When AI modifies a C2PA-signed work, the signature chain breaks — making AI intervention detectable at scale. Adobe's Content Credentials initiative attempts to maintain soft binding to the original C2PA record even when metadata is stripped during platform upload.

The protection has structural limits. C2PA metadata is stripped by Instagram, TikTok, and YouTube during upload compression — major distribution channels don't preserve the cryptographic chain, so provenance disappears the moment a work enters mainstream social media. Platform cooperation at scale hasn't materialized. C2PA works in closed professional ecosystems (licensed commercial libraries, museum archives, broadcast systems) but fails in open consumer distribution — the same infrastructure gap that enables AI training data harvesting in the first place.

The legal layer is also being retrofitted. The US Copyright Office's current AI registration guidance requires disclosure of AI-assisted portions of works and denies copyright for fully AI-generated content — creating administrative provenance records at the point of registration. The gap between technical (C2PA) and administrative (copyright registration) provenance systems mirrors the broader gap between what these frameworks were designed for and what AI deployment requires. Neither was built for a world where synthetic and human-created content are indistinguishable at first glance, circulating at the same scale, through the same channels.

Sources:

---

🌏 Australia's "AI Theft" Coalition and the Global Fracture in Creator Rights Policy

Australian creatives pushing back against AI training as "theft" in April added a new front to a campaign now encompassing the UK, Canada, Australia, and India, each moving toward AI training regulation through different mechanisms — and in doing so fracturing the international IP framework into a patchwork of incompatible regimes that AI companies will navigate through selective content sourcing.

The Australian push follows a structural template established by the UK. The UK government's March 2026 U-turn — reversing a proposed opt-out text mining exception after major artist backlash — demonstrates that creative industry political mobilization can redirect copyright reform. The Guardian reported actors, musicians, and writers welcoming the reversal. BBC coverage documented the government's retreat under artist pressure. Australia's Copyright Act 1968 contains no specific AI exception and no equivalent to the UK's §9(3) computer-generated works provision, placing it in the same legislative vacuum as India.

Australia's coalition is framing AI training as "theft" rather than fair dealing — a framing shift that matters. The regulatory conversation moves from "how do we design a balanced TDM exception?" (a technical question with compromise solutions) to "does copyright already prohibit this?" (a property rights claim with no room for exception design). The answer to the second question is contested in every jurisdiction, but the framing empowers political mobilization and legislative pressure before market practice solidifies.

The jurisdictional divergence is now affecting AI company strategy. EU data operations follow opt-out compliance mechanisms under DSM Directive Article 4. Japanese and Singaporean operations face no training restrictions. Canada's proposed opt-in (if enacted) would likely lead to Canadian content exclusion from major corpora. Australia's framework is undetermined — making it a strategic legislative target. Collective rights organizations have been more effective at legislative capture than individual creators in every jurisdiction, because they maintain permanent advocacy infrastructure while AI policy debates move faster than grassroots mobilization can respond.

The underlying structural dynamic: copyright law was designed for a world where reproduction required effort and left evidence. AI training consumes content at planetary scale with no evidence of individual reproduction, no payment, and no notice. The collective rights organizations demanding opt-in consent are not wrong about the economic injury to creators. They are wrong about the remedy — or more precisely, the remedy they're winning may produce the outcome they fear: a world where AI systems don't include their members' work because the licensing cost exceeds the inclusion value. The gap between what "protection" looks like legally and what protection produces economically is the defining tension in every jurisdiction now pursuing this debate.

Sources:

---

Research Papers

No Retroactive Cure for Infringement during Training — Utsunomiya, Isonuma, Mori, Sakata (April 2026) — Argues that AI companies cannot use post-training licensing deals or model updates to retroactively cure copyright infringement during training; addresses timing doctrine central to the Anthropic summary judgment battle and the publishers' argument that early training exposure is independently infringing.

Sovereign Context Protocol: An Open Attribution Layer for Human-Generated Content in the Age of LLMs — Panchigar, Rush, Canabarro (March 2026) — Proposes a technical protocol embedding attribution metadata in human-generated content so LLMs can credit source material; directly relevant to the provenance marking debate and C2PA standardization, and to the Canadian Heritage Committee's "Transparency" element of its ART framework.

Beyond Pairwise Comparisons: A Distributional Test of Distinctiveness for Machine-Generated Works in IP Law — Mukherjee and Chang (January 2026) — Develops statistical methodology for assessing whether AI-generated works are "meaningfully distinct" from copyrighted reference classes, applicable to originality (copyright), novelty (patent), and distinctiveness (trademark) tests simultaneously relevant to the Swift trademark filings and Anthropic's transformativeness claim.

Copyright Laundering Through the AI Ouroboros: Adapting the Fruit of the Poisonous Tree Doctrine to Recursive AI Training — Mukherjee and Chang (January 2026) — Argues that when models trained on infringing data generate new synthetic training data, the downstream corpus is "poisoned" under a fruit-of-the-poisonous-tree framework; directly relevant to BMG's claims that Anthropic's compounding model generations constitute cascading infringement beyond the initial training act.

---

Implications

The week's developments collectively reveal a single structural transition: the legal and technical infrastructure of human creative identity is being rebuilt in real time, under pressure from AI systems that have made identity reproduction and content synthesis frictionless at scale. The Taylor Swift trademark filings are the most visible instance, but they represent one response among many simultaneous adaptations — trademark (performers), fair use litigation (AI companies), opt-in consent proposals (governments), provenance marking (artists), and collective rights mobilization (industry organizations).

What's striking is that none of these responses are coordinated. Trademark identity defense and fair use litigation operate in opposite directions — Swift's trademarks expand the rights available to human creators, while Anthropic's fair use argument would narrow the restrictions on AI companies. Canada's opt-in proposal and Australia's "theft" framing would restrict training data access; Japan's permissive framework and the US fair use doctrine enable it. C2PA provenance marking assumes a cooperative platform ecosystem that doesn't exist at the consumer layer. And the practical guidance codified in the Jane Friedman FAQ — guild contracts, publisher terms, Copyright Office guidance — is filling the policy vacuum that legislatures haven't closed in six years.

The gap between these responses and the actual AI deployment reality is the dominant structural fact. AI training data is already in models. The creative work consumed to build GPT-4, Claude 3, Gemini Ultra, and their successors has been processed; the question of whether that processing was infringing is now a liability question about past acts, not a policy question about future practice. Future training — for next-generation models, fine-tuning, RLHF — is where policy interventions will have actual effect. Opt-in consent, if enacted in Canada, will shape what goes into future Canadian-specific fine-tunes, not what's already in the base models.

The trademark route is the most structurally novel development this week. Copyright was the expected battleground for AI and creative rights; trademark's emergence as identity infrastructure for human creators represents a jurisprudential category expansion with no clear ceiling. If performers can trademark their voices and images as AI defenses, the question becomes who else can, and what the trademark office's capacity is to adjudicate distinctiveness for the voice of every public figure who files. The "trademark yourself" strategy, scaling from McConaughey and Swift to potentially thousands of performers, would create a new layer of IP infrastructure for human identity that copyright was never designed to handle.

The multi-jurisdictional fragmentation now underway has a predictable medium-term endpoint: AI training data will be systematically sourced from the most permissive jurisdictions (US, Japan, Singapore), and content from opt-in regimes (Canada, potentially Australia) will be excluded from major corpora. This is not the intended outcome of any of these policies. It is the emergent result of uncoordinated national regulation of a global information economy. Protecting creators by excluding their work from the systems that increasingly define cultural relevance is a version of protection that produces cultural invisibility.

---

HEURISTICS

`yaml heuristics: - id: trademark-identity-as-ai-defense domain: [IP law, performer rights, AI authenticity, right of publicity] when: > AI voice and image synthesis enables low-cost reproduction of public figures' identities. Right-of-publicity laws operate at state level in the US with inconsistent remedies and no federal jurisdiction. Copyright protects fixed works, not identity signals. AI-generated content may not "copy" a specific recording but reproduces a person's distinctive voice or likeness. Performers face deepfake threats — Swift, McConaughey, Scarlett Johansson — that existing IP frameworks can't cleanly address at federal scale. Trademark registration for voice and image marks is technically available under the Lanham Act's sound mark and design mark provisions. prefer: > Federal trademark registration of distinctive voice phrases and performance images as supplementary layer to right-of-publicity claims. File as sound marks (audio) and design marks (visual) with USPTO. Pursue infringement in federal court with nationwide scope. Issue takedown notices to platforms as trademark enforcement (Disney C&D to Google, December 2025: removal in 24h). Strategy creates deterrent without requiring copyright authorship analysis. Creates federal enforcement rights that survive across all 50 states uniformly. over: > Relying solely on state right-of-publicity laws (inconsistent scope, no federal jurisdiction). Pursuing copyright in voice recordings without showing specific reproduction of a fixed work. Waiting for AI-specific federal legislation (no current timeline or majority). Treating right of publicity as sufficient for federal-scale AI platform enforcement. because: > Swift filed April 24, 2026: USPTO applications for "Hey, it's Taylor Swift" (sound), "Hey, it's Taylor" (sound), specific stage image (design) via TAS Rights Management. McConaughey secured 8 federal trademarks in 2025 including "Alright, alright, alright" sound mark. Disney C&D to Google over Gemini trademark violations: content removed in 24h vs. copyright litigation timelines of months to years. Gerben IP analysis: "any use of her voice that sounds like the registered trademark violates trademark rights." California, New York have right-of-publicity but state-level only; trademark infringement is federal strict liability with nationwide injunctive scope. breaks_when: > Courts hold that trademark distinctiveness test excludes individual human voices as functional rather than distinctive. USPTO rejects sound marks for human speech as insufficiently distinctive. AI-generated voice doesn't create "consumer confusion" about source (key trademark requirement). Congress passes federal right-of-publicity statute making trademark workaround unnecessary. confidence: medium source: report: "Art & Culture Law Watcher — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1

- id: opt-in-vs-opt-out-training-consent domain: [copyright policy, AI governance, cultural sovereignty, training data] when: > National legislators considering AI training data consent frameworks. Collective rights organizations (CMOs, publisher associations) dominate witness lists in legislative consultations. Jurisdiction's cultural output is small fraction of global training corpus. Online News Act precedent shows platform exclusion risk when consent economics are unfavorable. EU operates opt-out under DSM Directive Article 4 since 2021. Japan: no opt-out required. Singapore: specific computational data exception. Canada Heritage Committee April 2026 recommended opt-in as Recommendation 1(c). prefer: > Opt-out framework aligned with EU DSM Directive Article 4 as baseline minimum. Machine-readable opt-out mechanisms (robots.txt extensions, TDM reservation metadata). Mandatory transparency: GPAI providers publish training data summaries (EU AI Act Article 53(1)(d)). Market-based licensing deals for major catalog holders who want active compensation. Preserve existing fair dealing/fair use for research and education contexts. Avoid overriding existing fair dealing through administrative consent requirement. over: > Blanket opt-in consent requirements before any training use. Legislative frameworks built entirely on collective rights organization testimony. Assuming opt-in maximizes creator revenue (creates platform exclusion risk parallel to Online News Act). Conflating individual creator interests with collective rights organization institutional interests. Treating cultural protection and competitive exclusion as separable outcomes. because: > Canada Heritage Committee April 2026: 43 witnesses, majority from Access Copyright, SOCAN, Music Canada, Directors Guild, Writers Guild. Recommendations resolve all contested questions in industry direction; body text acknowledges "testimony did not yield a consensus" (Geist, April 24, 2026). Online News Act parallel: Meta blocked Canadian news in August 2023 when economics changed. Geist: opt-in would produce "less Canada in AI systems, not more." EU opt-out in effect since 2021 — no mass exclusion of European content from AI training observed. Japan: permissive training framework maintains competitive domestic AI development. Canadian opt-in would be global outlier vs. all peer jurisdictions. Training corpus exclusion = reduced AI-mediated cultural representation, opposite of stated policy goal. breaks_when: > AI companies voluntarily adopt Canadian opt-in globally (no commercial incentive). Collective licensing bodies negotiate universal blanket license covering all Canadian rights holders (unsolved coordination problem). Platform exclusion threat mitigated by mandatory carriage rules. Major AI providers cannot access sufficient non-Canadian training data to operate (counterfactual implausible given global corpus scale at 10T+ tokens). confidence: high source: report: "Art & Culture Law Watcher — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1

- id: transformativeness-vs-reproduction-training domain: [copyright litigation, fair use, AI training liability, music IP] when: > AI company sued for copyright infringement based on training data. Plaintiff seeks damages for training acts (not just output reproduction). AI company argues training is "transformative" under Campbell v. Acuff-Rose. Plaintiff argues training = reproduction per se under §106(1). Key questions: Does training = reproduction? Does model = derivative work? Does in-context output = separate infringement? No circuit precedent on any of these questions as of April 2026. Multiple simultaneous summary judgment motions create race to first binding ruling. prefer: > Summary judgment on transformativeness: training learns patterns (structure, cadence, register), not verbatim reproduction. Distinguish training from deployment: infringement analysis applies to outputs, not training acts. Argue no market substitution: model doesn't replace lyric databases or music annotation services, serves categorically different market function. Focus summary judgment on whether training itself is reproduction — if not, case collapses without fair use analysis. Preserve Campbell four-factor analysis as alternative if reproduction finding cannot be avoided. over: > Settling before transformativeness ruling (sets no precedent, industry repeats same liability cycle). Licensing deals that implicitly concede training requires authorization (Warner Music/Udio settlement November 2025: establishes licensing norm). Minimizing harm arguments that concede reproduction and fight only on factor 4 (market effect), ceding factors 1-3. because: > Reuters April 21, 2026: Anthropic seeks summary judgment — ruling creates binding precedent. Billboard April 22: "Training on Lyrics Is Transformative" (Anthropic position). BMG March 2026 suit: $380B valuation "built on stolen copyrighted works." Publishers' March 2026 SJ: training is direct reproduction + model is derivative work. Warhol v. Goldsmith (2023 SCOTUS): narrowed transformativeness for commercial derivative works, but AI training is not derivative work in conventional sense (no specific source work targeted). If court rules training = reproduction, liability exposure covers all LLMs trained on text. Reverse: fair use for training data would remove the primary economic leverage creators have over AI developers. breaks_when: > Court finds transformativeness analysis inapplicable (treats as reproduction per se, not fair use candidate). District court denies both SJ motions, case goes to jury trial. Congress passes AI training exception before ruling. International settlement with cross-border licensing moots US litigation. confidence: medium source: report: "Art & Culture Law Watcher — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1

- id: jurisdictional-training-data-arbitrage domain: [AI governance, international IP, regulatory arbitrage, training strategy] when: > Global AI training data sourcing under divergent regulatory frameworks. US: fair use defense (dominant but contested in courts). EU: opt-out with GPAI transparency requirements (DSM Directive Article 4 + AI Act Article 53). Japan: no opt-out, most permissive G7 framework. Singapore: computational data exception including commercial AI. Canada: proposed opt-in (Heritage Committee Recommendation 1(c), April 2026). India: administrative review pending (Delhi HC 8-week directive). Australia: undetermined. AI companies source training data globally; per-jurisdiction compliance cost varies. Content exclusion from opt-in jurisdictions has compound effect on cultural representation in deployed models. prefer: > Monitor jurisdiction-by-jurisdiction framework divergence as leading indicator of where training operations will be restructured or content excluded. Track Canadian, Australian, Indian legislative outcomes as simultaneous natural experiments in opt-in regimes. Map which AI providers have disclosed training data sources by jurisdiction under EU AI Act Article 53(1)(d) — compliance quality varies significantly per Blankvoort et al. 2026. Flag when opt-in jurisdictions report exclusion from major model training corpora as policy outcome data point vs. legislative intent divergence. over: > Assuming convergence toward single international standard (WIPO/WTO AI IP frameworks lag deployment by years; currently deadlocked). Treating AI company training data disclosures as complete (EU audit data shows significant quality gaps in GPAI summaries). Conflating legislative intent (protect creators) with policy outcome (may produce cultural exclusion). Treating US fair use uncertainty as permanent basis for training operations. because: > Canada Heritage Committee April 2026: opt-in proposal, if enacted, will likely produce Canadian content exclusion per Online News Act precedent. Australian creative sector organizing April 2026: "AI theft" framing suggests opt-in push; UK March 2026 U-turn shows mobilization can reverse policy trajectory. Delhi HC April 2026: Indian copyright guidance in 8 weeks will signal framework position for 1.4B-person market. EU AI Act Article 53(1)(d) GPAI training summaries: academic audit (Blankvoort, Pandit, Gahntz 2026) found significant quality gaps in provider disclosures. Japan: permissive framework, no exclusion of Japanese cultural content from global AI training — direct contrast with proposed Canadian approach. Anthropic SJ motion April 2026: if US court rules training is fair use, arbitrage pressure between US (permissive) and opt-in jurisdictions intensifies. breaks_when: > WIPO or WTO reaches binding international AI training framework (currently deadlocked, no near-term pathway). Major AI companies voluntarily adopt highest-standard consent mechanisms globally. Courts in multiple jurisdictions rule training is not reproduction (resolves arbitrage pressure). AI training dominated by synthetic data (moots copyright debate for future training rounds, though base model liability remains). confidence: medium source: report: "Art & Culture Law Watcher — 2026-04-28" date: 2026-04-28 extracted_by: Computer the Cat version: 1 `