The legal front around AI widened in 2026, moving beyond the original training-data copyright suits into discovery fights, answer-engine cases, and the first rulings on whether model weights are infringing copies. Nothing here is settled law yet, but the direction matters for anyone training on or building with third-party content. Note that court statuses change quickly; the items below reflect the record as researched and each should be checked against the live docket before relying on it.

What happened

The New York Times and other news plaintiffs filed a motion on 9 July 2026 asking the court to sanction OpenAI, alleging it deleted ChatGPT logs in violation of a preservation order (case 1:23-cv-11195, S.D.N.Y.). It is a filed motion, not a ruling, but it is the first major spoliation fight in the flagship AI-copyright case.

In Bartz v. Anthropic (N.D. Cal.), the roughly $1.5 billion settlement, which would be the largest US copyright settlement ever, had its final approval hearing on 14 May 2026 and was taken under submission; final approval had not been entered as of the latest available record. Meanwhile the litigation broadened: Encyclopaedia Britannica and Merriam-Webster sued OpenAI (13 March), BMG sued Anthropic (17 March), and CNN sued Perplexity (28 May), pushing the fight into reference works, music, and answer engines. In the UK, the Getty v. Stability AI judgment (November 2025) held that model weights are not “infringing copies” while finding limited trademark infringement, with an appeal window in early 2026. On the regulatory side, Ireland’s Data Protection Commission opened a GDPR inquiry into X and xAI on 17 February 2026 over non-consensual deepfakes generated by Grok.

Why it matters for builders

Two practical signals. First, provenance and retention are now legal exposure. The OpenAI sanctions fight is about deleted logs; if you build a product on top of an LLM, how you log, retain, and can produce records is becoming a discovery question. Second, the “is training fair use” question is still open, and it is now being tested across text, music, reference works, and answer engines at once, so a ruling in any one case could ripple.

The Getty UK holding, that weights are not copies, is an early data point favouring model builders on one narrow question, but it is under appeal and is UK law only. The safe posture is unchanged: know the provenance of your training and retrieval data, keep records, and treat “fair use will protect us” as an unsettled bet. Pair this with the regulatory compliance checklist .

Update, September 2026

Doe v. GitHub (Ninth Circuit, 16 September 2026). A unanimous panel upheld the dismissal of the open-source programmers’ DMCA §1202(b) claims against GitHub, Microsoft and OpenAI. It held that Copilot and Codex, as the complaint describes them, generate new code that never contained copyright management information, so no CMI was “removed or altered.” The court rejected the district court’s strict “identicality” rule, found the plaintiffs had standing, and expressly left open ordinary copyright-infringement claims. It did not decide a training-stage CMI theory, because that theory had not been preserved. Breach-of-contract claims remain pending in the trial court. Full write-up: Ninth Circuit upholds dismissal of DMCA claims in Doe v. GitHub .

China proposes a fair-use regime for AI training. China’s National Copyright Administration published its copyright development plan for 2026-2030, released online on 7 September 2026 and reported by Digital Watch on 9 September. Among other measures, it proposes developing a fair-use system for AI training data, which it describes as balancing innovation with protection of creators’ rights. The plan sets a direction rather than specific rules. If adopted, it would settle in policy a question that US courts are still deciding case by case.

Sources

Further reading