Fair Use and AI Training in US Law

Fair use is the defense in US copyright law that decides whether an AI developer may copy protected works to train a model without a license, and as of November 10, 2026 no statute or Supreme Court ruling answers that question for training. Two district courts held training fair use in June 2025 on their records, the Third Circuit held on September 29, 2026 that a non-generative tool's training was not, and three summary-judgment rulings were still pending as of November 9, 2026.

17 U.S.C. 107, in force since 1978-01-01, is the copyright layer of AI training data licensing in the United States. This page covers the four factors, how the copies were acquired, training versus outputs, dated rulings, government positions and statutory damages, for holders deciding whether to license and developers acquiring data.

Is AI Training Fair Use Under 17 U.S.C. 107?

It depends on the record: two federal district courts held training fair use in June 2025, and the Third Circuit held on September 29, 2026 that a non-generative legal search tool's training was not fair use. In Bartz v. Anthropic the order of 2025-06-23 (Dkt. 231) found training fair use. In Kadrey v. Meta the order of 2025-06-25 (Dkt. 598) granted summary judgment to Meta on that record and says it "does not stand for the proposition that Meta's use ... is lawful".

Fair use is decided per use and per record, so no ruling settles every training run: a different work, a different copy source or a different model can change the analysis. Three summary-judgment rulings were pending as of 2026-11-09: In re OpenAI Copyright Infringement Litigation (MDL No. 3143), Concord Music Group v. Anthropic and In re Mosaic LLM Litigation. Counts of AI copyright suits differ between trackers and none is official, so this page gives none.

The Four Fair Use Factors Applied to Training Copies

Section 107 directs a court to weigh four factors, and the training rulings so far turn mainly on the first and the fourth. Factors two and three appear below with the statutory wording only; this page does not summarize what the rulings say on them, and the table marks them so.

The table below sets each statutory factor against what the training rulings say, as of 2026-11-10.

Purpose and Character: Why Generative and Non-generative Training Diverge

The first factor asks whether the use is transformative and commercial, and the courts have answered it differently for a model that generates new text and for one that returns existing material. Bartz held training "exceedingly transformative" (Dkt. 231). The Third Circuit, in an opinion by Judge Montgomery-Reeves (No. 25-2153, 2026-09-29), held ROSS's use of 2,243 headnotes "minimally transformative, at best" and commercial, and noted that ROSS's AI "was not a generative AI". It is the first federal appellate fair use ruling on AI training.

The Statement of Interest of the United States, filed 2026-09-01 in MDL No. 3143 and known from its citation in footnote 7 of the Third Circuit opinion, argued that training which can generate original responses is transformative. It is a party's filing, not a ruling.

Market Effect: Lost License Fees, Market Dilution and Substitution

The fourth factor asks whether the use harms the market for the work, and courts disagree on whether a market for training licenses counts. Three theories appear in the record.

• Lost licensing market: Kadrey says the plaintiffs are "not entitled to the market for licensing their works as AI training data", while the Third Circuit says "the market for licensing headnotes as text to train AI is rapidly developing".
• Market dilution: Kadrey calls it the "potentially winning argument" but undeveloped on that record; in MDL No. 3143 the order of 2026-10-06 (Dkt. 2090) precluded class plaintiffs from relying on the Chakrabarty paper, granted without opposition.
• Substitution: the Statement of Interest argued that training causes no substitutive competition, a party's position.

The Concord court asked the parties about market dilution "in a but-for licensing world" (Dkt. 918), so the question stays open in a third pending case.

How the Training Copies Were Acquired: Purchased, Licensed or Pirated

How a developer obtained the works can decide the case even where training itself is fair use: in Bartz, scanning purchased print books was fair use, while the pirated library copies were not excused. Dkt. 231 counts the Books3 copying at 196,640 books (early 2021), at least five million LibGen copies (June 2021) and at least two million PiLiMi copies (July 2022), with over seven million pirated copies kept. The case then settled for USD 1.5 billion, with final approval and judgment on 2026-07-20 (Dkt. 680).

The Copyright Office's pre-publication Part 3 (p. 52) says knowing use of pirated or illegally accessed works weighs against fair use without being determinative. In Huckabee v. Bloomberg (S.D.N.Y. 1:23-cv-09152-MMG) the motion to dismiss was denied on 2025-11-24 (Dkt. 98): Books3 training was plausibly alleged and fair use was not resolved at the pleading stage. Licensed copies come with documented provenance, which answers the acquisition question directly, though no ruling says a license guarantees fair use for every use; the sources are covered in pirated books and shadow libraries in AI training.

Training Copies Versus Model Outputs in a Fair Use Defense

Courts treat the copies made to train a model and the content the model later produces as separate uses, each with its own fair use analysis. In Concord Music Group v. Anthropic (N.D. Cal. 5:24-cv-03811-EKL), pre-hearing questions published by Clerk's Notice on 2026-10-23 (Dkt. 918) split fair use four ways: training, verbatim lyric outputs, mash-up outputs and model-written new lyrics. The ruling was pending as of 2026-11-09.

A German court has applied US law to training. In GEMA v. Suno (LG München I, 42 O 763/25, 2026-07-31) the court held that training copies made in the US were governed by US law and were not fair use, with outputs substantially similar; the judgment is not final. Whether outputs can themselves be protected is outside this reference.

Fair Use Rulings and Pending Decisions on AI Training, Dated

Eight proceedings have produced a ruling that addresses fair use in AI training or have one pending, and each row gives the court, docket entry, outcome on fair use and status as of its check date (2026-11-09 unless stated).

Only the Third Circuit ruling is appellate. ROSS petitioned for panel rehearing or rehearing en banc on 2026-11-03 (Doc. 225), and no Supreme Court step is stated.

Claims, settlements and non-US cases are tracked in the tracker of AI copyright lawsuits over training data.

What the US Government Has Said About Fair Use for Training

Two federal bodies have taken positions on fair use for training, and neither binds a court: the Copyright Office in a pre-publication report and the United States in a Statement of Interest. The positions are dated and kept apart from law.

• Copyright Office Part 3, pre-publication version of 2025-05-09: the final version was not published as of 2026-11-09, and copyright.gov says a final will follow without substantive changes.
• Pirated data (p. 52): knowing use weighs against fair use, not determinatively.
• Licensing (p. 106): the Office would let the licensing market develop without intervention, with extended collective licensing only where market failure is shown.
• Statement of Interest, 2026-09-01: training that can generate original responses is transformative, a position in one case.

The Office's report has its own page, the US Copyright Office report on generative AI training.

Statutory Damages When a Fair Use Defense Fails

Where fair use fails, statutory damages under 17 U.S.C. 504(c), in force since 1978-01-01 and checked 2026-11-10, run from USD 750 to USD 30,000 per work, up to USD 150,000 if the infringement is willful and down to USD 200 if innocent. Per-work amounts matter because training sets hold thousands to millions of works, so even the lower bounds multiply quickly, though no exposure figure is computed here because the number of works at issue in any set is a matter of proof.

The Bartz notice and order cite these figures, and the settlement closed the pirated-copies claim before trial; the per-work payments belong to the case page. Statutory damages arise only after a court rejects fair use, so a ruling for the developer ends that exposure on the records decided so far.

Licensing Works or Data Under US Fair Use: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

US fair use does not make a license unnecessary or necessary in general, so a holder weighs what licensing adds against the current rulings, and not licensing remains a valid choice.

Benefits

• Lawful, documented copies are the route Bartz treated as fair use and Part 3 favors over pirated sources.
• A developing licensing market weighed in ROSS.
• A license fixes scope and price by contract, independent of how pending cases end; a rights audit before licensing data confirms the holder owns what it licenses.

Risks

• Kadrey refused a licensing-market claim on its record.
• Training held fair use in Bartz and Kadrey lowers the leverage of an unlicensed holder.
• Statutory damages apply only after a court rejects fair use.

Downsides

• Rulings are pending in MDL No. 3143, Concord, Mosaic and the ROSS rehearing petition.
• None of the rulings above sets a price for a training license.

Alternatives

• Not licensing at all.
• Licensing only works the holder owns.
• Collective licensing.
• Waiting for the pending rulings, with no outcome implied.

The full decision is set out in should you sell or license your data to AI companies.

Beyond Fair Use: Other US Rules, Other Jurisdictions and the Seller Process

Fair use is one US copyright rule, and other countries use different exceptions. Privacy, contract and state rules for a data deal sit on the United States AI training data law page, and the list below leads to the rest.

• Copyright and AI training across jurisdictions: exceptions outside the United States.
• AI training data law by jurisdiction: every jurisdiction.
• Is web scraping for AI training legal: contract and computer-misuse claims beside copyright.
• AI data licensing agreements: warranty and indemnity clauses on copyright.
• News publishers licensing content to AI companies: publishers who license rather than litigate.
• Federal AI training data bills: bills that would change disclosure, not fair use.
• Books as AI training data: book licensing and prices.
• Collective licensing for AI training: licensing through a collective.
• Alternatives to selling data to AI labs: alternatives, including not selling.

What This Page Does Not Cover

This page covers copies made to train AI models, not whether AI outputs can be protected by copyright or reused, such as song covers made with AI. Usage limits of AI tools are outside the reference too.

How This Page Is Sourced and Dated

Every ruling here cites the court, docket entry and date, and every status is given as of the date it was checked. This page is general information about US copyright law, not legal advice for any specific deal or dispute.

Frequently Asked Questions

Has a court ruled on whether OpenAI's training is fair use?

No: as of November 9, 2026 no summary-judgment ruling on fair use had been located in MDL No. 3143. OpenAI filed its opposition on 2026-10-29 (Dkt. 2312), and the United States filed a Statement of Interest on 2026-09-01.

Did Meta win its fair use defense over AI training?

On its record, yes: Kadrey v. Meta granted Meta summary judgment on training on June 25, 2025. The order "does not stand for the proposition" that Meta's use is lawful, and a related Entrepreneur trial is set for 2027-05-03.

Does US fair use apply to training done outside the United States?

A court outside the United States may apply it. In GEMA v. Suno (2026-07-31) the Munich court applied US law to training copies made in the US and found them not fair use; the judgment is not final.