What the Anthropic settlement did not decide
Final approval landed 20 July 2026, but a settlement creates no precedent. The ruling underneath it split lawful acquisition from training.
The number everyone repeats is $1.5 billion. The number that matters more is zero, which is the amount of binding precedent the settlement created.
Final approval in the Bartz v. Anthropic class settlement landed on 20 July 2026. Since then it has been cited in pitch decks, LinkedIn posts and at least one vendor security questionnaire as though a court had ruled on AI training. No court ruled on anything in July. A court approved a deal.
The ruling that actually exists came a year earlier, says something quite specific, and says almost none of what people are attributing to it.
Informational, not legal advice.
What was approved
A class settlement of $1.5 billion, given final approval on 20 July 2026 by Judge Martinez-Olguín, who took the case over after Judge Alsup went inactive. The Authors Guild's summary for authors is the clearest walkthrough of the claims process.
What a settlement is, mechanically: an agreement to stop litigating in exchange for money, approved by a court as fair to the class. It resolves the claims of those class members. It does not decide the legal questions the claims raised, and it does not bind any other court, any other defendant, or Anthropic itself in a future case with a different class.
Nobody involved is pretending otherwise. It is the coverage that drifted.
The ruling underneath it
The substantive decision came from the Northern District of California on 23 June 2025, and it split the conduct in two.
Training on lawfully acquired books was fair use. The court described the use as spectacularly transformative. Buying books, digitising them, and training a language model on them was held to be a paradigmatically transformative use.
Retaining a pirated library was not. Downloading and keeping a library of pirated books was a separate act, and it did not become fair use because of what the books were later used for. That is what the class action was about, and that is what the $1.5 billion resolved.
The distinction is the whole story and it is the part that gets flattened in retellings. The liability that generated the largest sum paid in this field so far attached to acquisition, not to training. Anthropic won the training question at district level and paid for the sourcing question.
For anyone assessing model suppliers, that reframes the diligence question. "What did you train on" is less operative than "how did you obtain it", and the second question is the one that has produced a nine-figure consequence.
Four things it does not decide
1. It is not appellate precedent, or any precedent. A district court ruling binds nobody outside that case, and a settlement binds even less. There is still no US appellate ruling on AI-training fair use.
2. It does not settle training generally. The June 2025 ruling concerned books and a language model. It does not carry across to images, video, music or code by its own force, and it did not purport to.
3. It does not clear Anthropic of everything. The lyrics litigation brought by Concord, UMPG and ABKCO involves two live suits with no settlement. Different works, different claims, unresolved.
4. It does not make the regulator's position law either. The US Copyright Office's Part 3 report of 9 May 2025 rejects the claim that training is inherently transformative. That report remains a pre-publication draft with no final version issued as of August 2026, and a Copyright Office report is not law regardless of its status. It is a well-informed argument, and courts are free to disagree with it, which the June 2025 ruling effectively did.
The scoreboard, as it actually stands
| Matter | Status | What it establishes |
|---|---|---|
| Bartz v. Anthropic (N.D. Cal.) | Ruling 23 Jun 2025; settlement final 20 Jul 2026 | Lawful acquisition + training = fair use; pirated retention = not. District level only. |
| Kadrey v. Meta (N.D. Cal.) | Summary judgment for Meta 25 Jun 2025; interlocutory appeal denied Jul 2026 | Decided expressly on a thin record. Torrenting phase unresolved, hearing 25 Feb 2027. |
| Thomson Reuters v. Ross (D. Del.) | Not fair use, 11 Feb 2025; Third Circuit argued 11 Jun 2026 | Non-generative tool, direct market substitute. Decision pending — first appellate word. |
| Getty v. Stability (EWHC) | Judgment 4 Nov 2025 | Training and output claims dropped mid-trial. UK training legality never decided. |
| NYT v. OpenAI/Microsoft (S.D.N.Y.) | Discovery; sanctions motion filed Jul 2026 | Nothing yet. No trial date. |
| Disney, Universal, Warner Bros. v. Midjourney (C.D. Cal.) | Consolidated 4 Nov 2025; expert disclosures Oct 2026 | Nothing yet. Targets outputs, not just training. |
Two district rulings pointing opposite ways on adjacent facts, one appellate decision pending, and the largest sum so far paid for how material was obtained. That is the entire body of US authority. Anyone describing this as settled in either direction is selling something.
The Midjourney row deserves a second look, because it is the one closest to a working studio's actual exposure. That case is about outputs depicting protected characters, not only about what went into training. Provider terms that assign you the output do not immunise you against a third-party infringement claim over what the output depicts, and that gap is unresolved. Legal and licensing for AI content in business covers where the user-side line sits today.
Music is on a different track entirely
The music settlements are frequently folded into the same narrative and should not be. UMG settled with Udio in October 2025; WMG settled with Udio and with Suno in November 2025. Each pairs a litigation settlement with a forward licence covering opt-in artist catalogues.
What remains open: Sony Music has settled with neither Suno nor Udio, UMG v. Suno continues with Suno defending on fair use, and in June 2026 the AFM sued UMG and WMG over how settlement proceeds are allocated to session musicians. No merits ruling on AI music training exists. A licensed catalogue is not a cleared model, and why a label deal is not a cleared model is the version of this argument worth reading before you brief a music-heavy campaign. On the output side, the six-point chart test for AI music is the practical filter.
What a studio should actually change
Very little in your production process, and one thing in your buying process.
- Nothing about your ownership position. These cases are about inputs to training. Your rights in what you generate still run through human authorship and through your contract, unaffected by any of the above.
- Sourcing beats training claims in diligence. Ask providers how material was acquired, not just what it consisted of. That is where the money has actually gone. Licensed training data as a buying criterion is the version of this you can put in a supplier questionnaire.
- Keep the per-asset record. If a model is withdrawn, restricted or re-licensed mid-campaign, the studios that can say precisely which shots came from it reshoot a list. The ones that cannot reshoot the campaign.
- Prefer breadth over dependence. A production plan that only works if one specific model stays available is a legal risk dressed as a creative preference. A broad model catalogue is cheap insurance, and open-weight options come with their own conditions worth reading, which open-weight video licences covers.
FAQ
Does the settlement mean training on copyrighted books is now legal?
No. The June 2025 ruling held that training on lawfully acquired books was fair use, at district-court level, in one case. That is persuasive, not binding, and the settlement added nothing to it. The separate holding that retaining a pirated library was not fair use is the part that produced the payment.
Can I rely on the ruling for image, video or music models?
Not directly. The decision concerned a text model trained on books. Different media raise different fourth-factor market arguments, and the pending studio and music cases are being litigated on facts the books case never touched.
When will there actually be a binding answer?
The Third Circuit heard argument in Thomson Reuters v. Ross on 11 June 2026 and the decision is pending. That will be the first US appellate ruling on AI-training fair use. Note that it concerns a non-generative legal research tool held to be a direct market substitute, so it may sit awkwardly against generative facts even once it lands.
Should any of this change what I ship this quarter?
Only if you are shipping outputs that depict protected characters, real people, or recognisable brand assets, in which case the exposure was always output-side and always yours. The training cases do not touch it. The AI copyright and safety guide covers the checks that do.