Getty Defeat and Meta Fair Use Win Signal Shift in AI Copyright Battles

Two landmark rulings one in London, one in San Francisco have reshaped the immediate legal landscape for artificial intelligence developers and rights holders. The High Court in the UK has rejected the “model-as-copy” theory, finding that trained model weights do not themselves constitute infringing copies. In California, judges have extended fair use protection to AI training on lawfully acquired books, while warning that market-harm arguments remain alive. For lawyers and policymakers, the transatlantic picture is now one of narrow technical comfort for AI firms and open questions for creators.

Key Holding — Getty v Stability AI (UK):
“An AI model … which does not store or reproduce any Copyright Works (and has never done so) is not an ‘infringing copy’.”
— Mrs Justice Joanna Smith

Mrs Justice Smith’s reasoning was tightly drawn but highly consequential. The High Court held that, for secondary infringement under the CDPA, an “infringing copy” must embody a material reproduction of a protected work. Getty’s claim that the trained Stable Diffusion model contained infringing images collapsed because, in its final state, the model’s parameters neither stored nor reproduced any copyright works. The ruling turned purely on what the model is, not how it was trained. Primary infringement was not reached: Getty had withdrawn those allegations once it became clear that training did not occur in the UK.

The court did, however, uphold a limited trademark claim where AI-generated images displayed the Getty watermark. That was a narrow finding of unauthorised brand use, not a statement on copyright law. The broader issue — whether the acts of crawling, copying, and processing millions of images to train an AI system constitute primary infringement — remains legally unanswered in Britain.

Across the Atlantic, the same week delivered two fair-use victories for AI developers. In Kadrey v Meta, Judge Vince Chhabria ruled that training Meta’s Llama models on large corpora of books qualified as fair use, emphasising that the decision rested on the plaintiffs’ poor pleading rather than on any blanket approval of AI training. He explicitly warned that “market dilution” the prospect of generative systems flooding the market with near-substitutable works — could yet prove a winning argument in future litigation.

Days earlier, Judge Alsup in San Francisco held that Anthropic’s scanning of lawfully purchased physical books amounted to a “mere format change” and was fair use, but drew a bright line against maintaining any central repository of pirated digital texts. The distinction is practical: provenance and lawfulness of acquisition now define the boundary of permissible training in US law.

The litigation frontier has shifted from metaphysics — whether weights are copies — to logistics: where, under what licence, and with what market impact the training occurs.

For AI companies, the combined rulings provide a defensive map. They can point to the UK decision as authority that model weights are not copies and to the US cases as confirmation that transformative use on lawful data may be fair use. The strategy is already visible: geofenced training outside the UK, contractual provenance attestations from data suppliers, and watermark filters to prevent brand contamination. The legal safe harbour, however, extends only to the trained artefact, not to the process that produced it.

For creators and publishers, the task is evidential and legislative. Without proof of domestic copying, courts will not reach the primary infringement question. That may drive renewed lobbying for a UK statutory framework akin to the EU’s text-and-data-mining regime, which allows machine learning on protected works unless the rightsholder opts out. The IPO’s current consultation will determine whether Britain codifies that approach or leaves it to piecemeal litigation.

• Nexperia and the Law of Overreach — A legal analysis of state power, jurisdiction and proportionality in the Dutch move against Nexperia.

• The Scholar State in Global Competition: Wang Yi, Chinese Diplomacy and the Civilisational Divide — By Jaffa Levy, 2 November 2025, for Telegraph Online.

• Getty Images (US), Inc v Stability AI Ltd [2025] EWHC (Ch) — High Court judgment of Mrs Justice Joanna Smith, including findings on secondary infringement, model weights, and trade mark misuse.
• Kadrey et al. v Meta Platforms Inc, United States District Court, Northern District of California, Judge Vince Chhabria (June 2025) — ruling on fair use in training Meta’s Llama models using copyrighted texts.
• Authors Guild et al. v Anthropic PBC, United States District Court, Northern District of California, Judge William Alsup (June 2025) — judgment distinguishing fair use in scanning lawfully purchased physical books from use of pirated digital corpora.
• UK Intellectual Property Office (IPO) Consultation Papers on Text and Data Mining (2023–2025) — policy background on lawful access, fair dealing, and opt-out regimes.
• Directive (EU) 2019/790 on Copyright in the Digital Single Market (Articles 3–4) — EU framework governing text-and-data-mining exceptions for research and commercial use.
• Copyright, Designs and Patents Act 1988 (UK) — statutory basis for primary and secondary infringement and fair-dealing exceptions.
• Legal Services Act 2007 (UK) — definitions of reserved legal activities and “holding-out” provisions relevant to unauthorised legal practice.

This analysis draws exclusively on public court records, statutory instruments, and official consultations. No proprietary or third-party journalism is cited.