Will the US Supreme Court rule that training generative AI on copyrighted works without a license is not fair use by the end of 2030?

TL;DR

I estimate a 21% chance of a qualifying Supreme Court ruling by December 31, 2030. The main constraint is timing, not salience: no generative-AI training fair-use issue has reached a federal appellate merits decision yet, and the safest last Supreme Court term for a decision before the deadline is October Term 2029. If the Court reaches the merits in time, a rightsholder win is close to even, but the most likely overall outcome is still no qualifying Supreme Court holding by the deadline.

Context

The question is narrow. It does not ask whether AI outputs infringe, whether pirate-source acquisition is lawful, whether AI firms owe damages, or whether lower courts reject fair use. It resolves YES only if the Supreme Court itself issues a merits holding that the unlicensed generative-AI training before it is not fair use under 17 U.S.C. §107.

As of July 21, 2026, the lower-court record is active but immature. The Copyright Office released its Part 3 generative-AI training report in May 2025 after receiving more than 10,000 comments, and said a final version was expected without substantive changes to its analysis (U.S. Copyright Office AI page). The Office framed fair use as fact-specific: research or constrained non-substitutive uses are safer, while training systems to generate similar expressive outputs is weaker (Copyright Office Part 3 report).

Evidence

The historical base rate is low. The modern Supreme Court fair-use merits line is short: Sony in 1984, Harper & Row in 1985, Stewart in 1990, Campbell in 1994, Google v. Oracle in 2021, and Warhol in 2023. That is 6 major fair-use merits cases in about 40 years, or roughly one every 6–7 years, with mixed direction (Copyright Office Fair Use Index, Google v. Oracle, Warhol). The Court also has a severe intake filter: in October Term 2024, it had 3,856 cases docketed during the term, granted 70 cases for plenary review, and disposed of 64 cases by full opinions (Supreme Court Journal, OT2024 statistics).

The leading district decisions do not yet produce a clean Supreme Court vehicle. In Bartz v. Anthropic, Judge Alsup held on June 23, 2025 that using books to train Claude was fair use, but that Anthropic’s downloaded pirated copies used for a central library were not justified by fair use (Bartz fair-use order). The case was then resolved through a $1.5 billion settlement approved on July 20, 2026, after the court had already described the training holding as fair use and the pirate-library theory as the settlement driver (Reuters via UOL, July 20, 2026, Bartz preliminary approval opinion). That removes the most famous early generative-AI training case from the normal appellate path.

Kadrey v. Meta cuts the other way but is still not a ready vehicle. Judge Chhabria granted Meta summary judgment on June 25, 2025 because the plaintiffs did not build a market-harm record, while stressing that the ruling did not establish that Meta’s Llama training was lawful in general (Kadrey order). The same opinion said plaintiffs will often win in better-record cases and named news articles as potentially stronger for market dilution than the books at issue there (Kadrey order). That makes Kadrey important doctrine-shaping evidence, but a weak immediate path to a clear Supreme Court anti-fair-use holding because it is record-bound and other claims remain procedurally messy.

The most advanced AI-training appeal is Thomson Reuters v. ROSS, but it probably does not qualify. The district court held on February 11, 2025 that ROSS’s use of Westlaw headnotes to train a competing AI legal-search tool was not fair use, and the Third Circuit appeal was filed on June 24, 2025 and argued on June 11, 2026 (Thomson Reuters v. ROSS appeal docket). The problem is definitional: the district court emphasized that only non-generative AI was before it, while this question requires training a generative model (Loeb & Loeb summary of ROSS). ROSS can influence the law but is unlikely to resolve this market unless the Supreme Court writes beyond the case facts.

The best YES paths are music and news. UMG v. Suno was filed on June 24, 2024 and remains active in the District of Massachusetts (CourtListener Suno docket). UMG v. Udio was filed the same day in the Southern District of New York and remains active (CourtListener Udio docket). The OpenAI copyright MDL was filed in the Southern District of New York on April 11, 2025 and remained active with a last filing on July 17, 2026 (CourtListener OpenAI MDL docket). These cases have stronger substitution and licensing-market facts than Bartz, but they still need district rulings, circuit decisions, cert, briefing, argument, and opinions before December 31, 2030.

The timing math is tight. Supreme Court Rule 13 gives a party 90 days after a federal appellate judgment to file a cert petition (Supreme Court Rule 13). Cases granted after January are usually argued the next term, and a case argued in October Term 2030 would usually be decided in 2031, outside this question’s deadline (SCOTUSblog on January grant timing). So the real last train is a circuit merits decision by roughly mid-to-late 2029, followed by a cert grant early enough for October Term 2029 and an opinion by June 2030.

My model is:

$$P(YES)=P(M) \times P(A\mid M)$$

where $M$ is a Supreme Court merits decision by the deadline that squarely reaches generative-AI training fair use, and $A$ is an anti-fair-use holding under the facts before the Court. I estimate $P(M)=41\%$. That reflects a high chance that at least one generative-AI training appeal is ready by 2029, discounted for settlement, record-bound rulings, remands, procedural holdings, and the Court’s small merits docket. I estimate $P(A\mid M)=52\%$. The cases most likely to get there in time are likely to have plaintiff-favorable facts, especially music or news, but Google v. Oracle, the district wins for Anthropic and Meta on training, and the Court’s ability to remand keep the conditional probability close to even. The product is 21%.

What's non-obvious

The obvious story is that AI copyright is huge, so the Supreme Court will decide it soon. The better read is that vehicle quality is the bottleneck. ROSS is early but non-generative. Bartz was famous but settled after a defendant-favorable training ruling. Kadrey contains the strongest judicial language against unlicensed training, yet Meta won because the plaintiffs did not prove market harm.

Licensing markets cut both ways. The Copyright Office treated licensing as increasingly relevant to factor four and discussed voluntary licensing in its May 2025 report (Copyright Office Part 3 report). That helps rightsholders if a case reaches the merits. It also raises settlement pressure, which can remove the best plaintiff cases before they make appellate law.

Limitations

The biggest uncertainty is procedural. A single clean Second Circuit or First Circuit ruling in 2028 could move this estimate up sharply. A wave of settlements, or a Supreme Court decision in ROSS that absorbs the Court’s interest without qualifying under this question, would move it down.

The second uncertainty is merits framing. A Supreme Court opinion could say pirate-source acquisition is unlawful, outputs infringe, DMCA claims survive, or market-harm evidence is inadequate without clearly holding that the training itself is not fair use. Those outcomes may be bad for AI companies but still resolve NO here.

The third uncertainty is legislation. Congress could create a licensing, transparency, opt-out, or safe-harbor regime before the key appeals mature. That would not itself answer this question, but it could moot or reshape the cases before the Supreme Court reaches them.