AI Copyright Liability in 2026: Key Cases and Settlements Reshaping Training, Fair Use, and Output Liability

Six major US copyright cases involving AI have now produced rulings or settlements that collectively reshape the liability landscape for AI model training and deployment:

1. Thaler v. Perlmutter (settled law). On March 2, 2026, the US Supreme Court denied certiorari, leaving in place the DC Circuit's holding that AI-generated works without human authorship are not copyrightable. This is now settled law in the US.

2. Thomson Reuters v. Ross Intelligence (on appeal). Summary judgment for Thomson Reuters — the court found Westlaw headnotes were original and protected, and that Ross's use to train its AI legal research tool was not fair use. Currently on appeal before the Third Circuit. This is the most significant fair-use-in-training ruling against an AI developer to date.

3. Bartz et al. v. Anthropic (near-final settlement). A proposed $1.5 billion class-action settlement covering authors whose books were used in training. As of the May 14 fairness hearing, the opt-in rate reached 92.77%, with an estimated payout of ~$3,000–$3,100 per work. Judge Araceli Martínez-Olguín held a 75-minute hearing and did not immediately approve the deal, requesting a supplemental brief from Anthropic by May 21 on late opt-outs. Final approval expected soon.

4. Kadrey et al. v. Meta Platforms (partial dismissal). The court granted partial dismissal based on fair use for LLM training. Claims related to alleged reproduction of pirated works during torrent "seeding" remain active in the Northern District of California.

5. In Re OpenAI Copyright Infringement Litigation (pending). The multi-district litigation alleging infringement through both training data use and ChatGPT outputs remains pending in the Southern District of New York.

6. Disney et al. v. Midjourney (pending). Major entertainment companies allege Midjourney copied copyrighted works to train its image-generation model. Pending in the Central District of California.

Emerging trends:

• Fair use is being tested asymmetrically. Thomson Reuters says training on proprietary data for a competing product is not fair use. Kadrey's partial dismissal says training on publicly available books may be. The doctrinal fault line is whether the trained model competes with the original work's market.
• Licensing is accelerating as a risk mitigant. The Anthropic settlement and broader market movement toward formal licensing agreements (as Norton Rose Fulbright notes) is becoming the preferred path to reduce litigation exposure.
• Output liability remains unresolved. The OpenAI and Disney cases will test whether model providers bear liability for infringing outputs generated by users — arguably the highest-stakes open question for enterprise deployers of generative AI.