The Thesis
The global consensus on AI and copyright broke almost the moment it was institutionally confirmed. On 11 November 2025, the Regional Court of Munich I held in GEMA v. OpenAI (Case No. 42 O 14139/24) that memorization — the storage of protected lyrics in a language model's statistical parameters — is a reproduction under § 16 UrhG and Article 2 InfoSoc. Five months earlier, in Bartz v. Anthropic (23 June 2025), the Northern District of California had held that training on copyrighted works was "quintessentially transformative" and therefore fair use. The same fact pattern produced opposite answers.
In the Thaler pattern, the US, EU, UK, and China refused copyright to AI-only output through different routes but reached the same place: surface differentiation / deep convergence. GEMA runs the other way. Every major court now agrees, on the surface, that AI training is copyright-relevant, yet on the same fact pattern the EU, the US, and Japan reach entirely different doctrinal answers: surface convergence / deep divergence. The exact inverse.
The divergence is not accidental. In Thaler, personality theory (Persönlichkeitstheorie) and incentive theory happened to converge: an AI has no personality and needs no incentive, so both denied AI authorship. In GEMA that agreement breaks. Personality theory treats training as the unauthorized exploitation of an author's externalized personality; incentive theory treats it as transformative use toward new creation. Confronted with the same facts, the two diverge as each stays faithful to its own logic. What Thaler concealed, GEMA exposes.
First, GEMA v. OpenAI moves global AI-copyright doctrine into a zone of divergence, and that divergence polarizes industry structure into an EU license-first model and a US fair-use balancing model. Second, the divergence emerged where personality theory and incentive theory had only accidentally agreed in Thaler, and it is now perpetuated through the polarization of industry itself.
What GEMA Decided — Five Holdings
Memorization is reproduction: storage in a model's statistical parameters falls within Article 2 InfoSoc's protection of reproduction "by any means and in any form."
Output is making available: that a user can obtain the lyrics through a simple prompt is itself an act under § 19a UrhG.
The § 44b TDM exception does not apply: it covers the analytical step but not permanent storage in the model, which breaches the deletion requirement.
Operator liability: OpenAI, not the user, is directly liable. The first EU decision to recognize this head-on.
No CJEU referral: an acte clair position, even with Like Company pending.
Four Doctrines
A. Strict, consent-based: Germany (GEMA) and the UK government, which abandoned its TDM-with-opt-out option on 18 March 2026. License required.
B. Fair-use balancing: the US. Bartz — training is fair use, pirated storage is separate; a $1.5 billion settlement, the largest in US copyright history.
C. Permissive, non-enjoyment use: Japan's Article 30-4, Singapore's § 244 — broad exceptions legislated in advance.
D. Output-controlled, moderate: China, Li v. Liu plus the Beijing Internet Court's guidance.
The EU's own question is unresolved: Like Company v. Google (C-250/25), CJEU Grand Chamber hearing 10 March 2026, judgment not before 2027, will confirm or reverse GEMA's position across the EU.
A Bimodal Market
If Thaler built a single market by defining a negative space (public-domain default), GEMA builds a bimodal market by defining a positive obligation (licensable training).
Layer 1, licensed AI: Suno–WMG, the UMG–Udio platform, fully licensed startups. EU-friendly.
Layer 2, fair-use-claim AI: OpenAI and Anthropic (US market only), Meta (the EU outlier).
The industry's license-over-litigation turn took hold in music — UMG–Udio (October 2025), WMG–Udio (19 November 2025), WMG–Suno (25 November 2025, an unlicensed model phased out before a 2026 licensed launch, $2.45 billion valuation). The second suit, UMG/Concord/ABKCO v. Anthropic (28 January 2026, 20,000+ songs, over $3 billion), is the largest non-class-action copyright case in US music publishing. EU AI Act Article 53 GPAI obligations move to enforcement on 2 August 2026, and the European Parliament Resolution (10 March 2026, Voss) recommends a rebuttable presumption, an EUIPO opt-out registry, and sector collective licensing — the Brussels Effect arriving in AI.
Five Tests
The permanence of the split turns on five tests: (1) the OpenAI v. GEMA appeal at the OLG München; (2) the CJEU's Like Company (C-250/25); (3) GEMA v. Suno (judgment expected 12 June 2026) plus Sony v. Suno/Udio (US, summer 2026); (4) the stability of US fair use (the Bartz fairness hearing was held 14 May 2026 and taken under submission, with final approval pending; Kadrey's market-dilution warning may yet bite); and (5) pressure for sui generis legislation (the EP Resolution and the CDSM review). The present trajectory points to strengthening of the strict camp — the EP Resolution, the UK government's status quo, and the OLG Hamburg LAION decision are all aligned that way.
The Takeaway
Thaler showed the strength of an accidental consensus; GEMA shows its contingency. Personality theory and incentive theory are systematically different doctrinal infrastructures that reach different conclusions in different domains. They agreed in Thaler only because that domain forced a clear negative from both; the domain GEMA addresses does not. Once industry splits into a bimodal structure and capital, infrastructure, and talent are invested into the split, reversing it becomes institutionally hard. The present divergence is at once a doctrinal decision and a deferral — the most accurate way to read the consensus that broke.