Prolepsia forecastYes / no

Will a US appeals court rule that training AI on copyrighted web content is fair use before 31 December 2027?

As of 17 November 2026, Prolepsia puts the chance of yes at 26%.

26%chance of yes

Yes 26%No 74%
Prolepsia forecast made

How Prolepsia reasons

The report Prolepsia wrote with this forecast.

How forecasts are made · Track record

TL;DR

The forecast is 26.5% Yes and 73.5% No for a qualifying ruling issued between November 17, 2026, and December 30, 2027. No is more likely because the leading cases face substantial procedural and scheduling barriers before an appellate court can affirmatively decide training fair use.

Key evidence

  • The strongest existing favorable ruling is not already on a clear appellate timetable. Kadrey’s district-court decision protected Meta’s training use on the particular evidentiary record, involving books downloaded from online sources (decision summary). But interlocutory review was denied on July 8, 2026, leaving further district proceedings before ordinary appellate review (docket).
  • Bartz is no longer a straightforward merits-appeal route. Final settlement approval on July 20, 2026, removes the ordinary expectation that its favorable training ruling will proceed to appellate review (approval order).
  • Other cases provide real opportunities, but remain at district-court milestones. Mosaic’s summary-judgment hearing is scheduled for December 18, 2026 (court calendar). OpenAI’s signed scheduling order sets summary-judgment replies for November 20, 2026 (order). Neither establishes an appellate merits schedule.
  • Historical appeal durations are unfavorable. For other civil appeals terminated on the merits during October 1, 2024–September 30, 2025, median notice-to-disposition times were 13.8 months in the Ninth Circuit and 11.7 months in the Second Circuit. Those intervals exclude the preceding wait for an appealable district decision (official statistics).

Uncertainties

An early final judgment, accepted interlocutory appeal, or expedited appellate schedule would create a stronger route to Yes. A narrow holding protecting training while condemning acquisition, retention, or outputs can qualify; the court need not approve every aspect of a developer’s conduct.

Settlement, surviving claims, remand, or resolution on unrelated grounds can prevent a qualifying decision. The substantive law also remains unsettled: the Third Circuit’s ROSS opinion rejected fair use for a non-generative system but expressly distinguished generative models (opinion). The available record does not establish a complete census of every potentially relevant appeal.

Read the full report

TL;DR

The forecast is 26.5% Yes and 73.5% No for a qualifying federal appellate ruling between November 17, 2026, and December 30, 2027. The decisive constraint is completing the appellate process before the deadline, not establishing that an affirmative fair-use argument exists. Watch for an appealable judgment and an appellate schedule, rather than another favorable district-court headline.

Context

The question concerns a specific judicial event, not the ultimate legality of AI training. A federal court of appeals must issue a published, precedential, controlling opinion expressly holding that at least one unauthorized use of internet-obtained copyrighted material for generative-model training is fair use. A narrow holding counts. A district-court ruling, an observation about transformativeness, or a remand without an affirmative fair-use determination does not. The last eligible opinion date is December 30, 2027.

The existing decisions establish a substantive opening but not a completed appellate route. Kadrey supplies a favorable, fact-specific district ruling (Copyright Office summary). Bartz settled after its training decision (final approval). The Third Circuit’s ROSS decision concerns non-generative AI and expressly distinguishes generative models (opinion). The remaining opportunity depends on live cases reaching an appellate holding in time.

Evidence

The historical backbone is appellate duration. There is no established empirical win-rate sample for this precise legal question in the available evidence. Broad civil-appeal statistics provide a firmer reference class for timing, though they do not isolate complex copyright cases or published precedential decisions.

The supplied historical benchmarks are below. They measure different case populations and reporting periods, so they are not one directly comparable trend.

Reporting period or vintage Population National median Second Circuit median Ninth Circuit median
October 1, 2024–September 30, 2025 Other civil appeals terminated on the merits 11.5 months 11.7 months 13.8 months
FY2025 annual report Merits dispositions — — 11.9 months
July 1, 2025–June 30, 2026 Broad merits-disposition benchmark — 15.9 months 11.5 months

The first row comes from the official Judicial Business table, with interval samples of 6,965 nationally, 760 in the Second Circuit, and 1,448 in the Ninth Circuit. The later profiles report 2,295 merits terminations in the Second Circuit and 5,175 in the Ninth Circuit (June 2026 profiles). The FY2025 Ninth Circuit annual-report figure is a separate benchmark.

These figures are not minimum waiting periods. They show why an appeal starting well into the forecast window faces a deadline problem. They also exclude the time needed to obtain an appealable district decision. The Ninth Circuit’s practice guide reinforces that constraint: ordinary non-priority civil appeals typically reach argument about 12 months after notice of appeal, generally four to five months after briefing (practice guide). Those intervals should not be added to the disposition medians; they describe different endpoints.

Appealability comes before speed. Ordinary appellate jurisdiction generally requires a final decision (28 U.S.C. § 1291). Interlocutory review under Section 1292(b) requires district-court certification and appellate permission (statute). A favorable training ruling can therefore leave the parties litigating distribution, acquisition, outputs, or other claims before an ordinary appeal begins.

Kadrey illustrates the distinction. The June 25, 2025 ruling protected Meta’s training use of books obtained from online shadow libraries, based on the record presented. The court emphasized shortcomings in the plaintiffs’ evidence of market harm, rather than announcing categorical protection for AI training (decision summary). An appellate court’s express adoption of that narrow determination could satisfy this question.

But Kadrey’s request for interlocutory review was denied on July 8, 2026. The reproduced court order favored reviewing the issues together after final judgment (docket). Meta’s third-quarter disclosure identifies remaining distribution and contributory claims and a February 25, 2027 summary-judgment hearing (disclosure). My reading is that the favorable training merits do not translate into an advanced appellate timetable.

Bartz presents a different barrier. Its district decision separated training from acquiring and retaining pirated library copies (Copyright Office summary). Final settlement approval followed on July 20, 2026 (approval order). That removes the straightforward route from the training ruling to an ordinary merits appeal. Settlement-related litigation is not itself an appellate determination of training fair use.

Other cases keep Yes credible. Concord’s questions for its October 21, 2026 summary-judgment hearing separate training from verbatim outputs, mash-ups, and synthetic lyrics. That framing creates room for a training holding distinct from output liability. It also raises the possibility that unresolved output claims delay finality. Judicial questions identify issues, not the eventual winner (court questions).

Mosaic provides another near-term district opportunity. The court moved its summary-judgment hearing from October 30 to December 18, 2026 (clerk’s notice). OpenAI’s signed order sets summary-judgment reply briefs for November 20, 2026 (scheduling order). Both are meaningful milestones. Neither is an appealable judgment or an appellate decision schedule.

The wider litigation field adds opportunities, but several schedules fit the deadline poorly. Andersen’s district summary-judgment and class-certification hearing is set for September 29, 2027, with trial unassigned (order). Midjourney’s earlier scheduling order provides a potential district motion route, but the current schedule and an actual training-summary-judgment filing were not established (scheduling order). These cases broaden the field without establishing a ready appellate vehicle.

Finally, ROSS does not settle the generative-AI issue. The September 29, 2026 Third Circuit opinion rejects fair use for a competing non-generative legal-research system and expressly distinguishes generative models (official opinion). It is neither a qualifying Yes nor a categorical bar. The Copyright Office’s training report likewise anticipates different outcomes across training practices, with attention to purpose, substitution, acquisition, and markets; it is analysis, not binding law (report).

What's non-obvious

A favorable district ruling is not an appellate countdown. The strongest substantive candidate can remain procedurally blocked by surviving claims. Kadrey’s denied interlocutory review makes that concrete. Conversely, a less prominent case that reaches a complete, appealable judgment earlier can matter more to this deadline than a better-known training victory.

The resolution threshold also cuts in the other direction. The court need not legalize all AI training, approve every dataset, or excuse unlawful acquisition. It needs to protect one challenged training use under the facts before it. Separating training from outputs and acquisition therefore has two effects: it creates room for a narrow qualifying holding, while leaving claims that can postpone appeal. More plaintiffs or copyrighted works do not automatically create more independent chances of a timely opinion when their cases share proceedings and legal issues.

Uncertainties

The largest uncertainty is the actual appellate timetable. The available record does not establish a complete census across all federal circuits or a verified accelerated merits schedule for the leading cases. That is a coverage limit, not proof that no relevant appeal exists. An accepted interlocutory appeal, an early final judgment followed by appeal, or a court-issued expedited schedule would close the most important gap.

The merits remain fact-sensitive. Kadrey demonstrates a favorable outcome on one record, not a general appellate rule. Stronger evidence of substitution or licensing-market harm could support a different result. Conversely, a narrowly framed training use could receive protection even if related conduct infringes. The decisive evidence would be the district court’s actual findings, the issues accepted for appeal, and the appellate court’s treatment of the statutory factors.

Settlement and procedural avoidance remain independent risks. A case can settle, proceed to further fact-finding, or be decided on unrelated grounds. Even a favorable opinion must be precedential, make the fair-use determination necessary to its disposition, and remain unwithdrawn and unvacated through resolution. These constraints explain why No remains the more likely outcome without implying that generative-model training will ultimately lose on fair use.

Forecast history

How Prolepsia’s forecast has movedOne forecast so far: 26% on 17 Nov 2026.17 Nov 2026: 26%
DateForecast
26% chance of yes

Sources Prolepsia read

  1. Kadrey v. Meta Platforms, Inc.www.copyright.gov
  2. Bartz et al v. Anthropic PBC, 4:2024cv05417 - Document 680 (N.D. Cal. 2026) :: Justialaw.justia.com
  3. opinionwww2.ca3.uscourts.gov
  4. https://www.uscourts.gov/sites/default/files/document/jb_na_app_0930.2025.pdfwww.uscourts.gov
  5. practice guidecdn.ca9.uscourts.gov
  6. 28 U.S. Code § 1291 - Final decisions of district courts | U.S. Code | US Law | LII / Legal Information Institutewww.law.cornell.edu
  7. statuteuscode.house.gov
  8. Cognella, Inc. v. Meta Platforms, Inc. 3:2026cv04053 | U.S. District Court for the Northern District of California | Justiadockets.justia.com
  9. 0001628280-26-066773d18rn0p25nwr6d.cloudfront.net
  10. Clerk’s notice publishing the court’s questions for the 21 Oct 2026 hearingchatgptiseatingtheworld.com

The question

As of November 17, 2026, US federal courts are actively considering whether using copyrighted material obtained from the internet to train artificial intelligence models without permission constitutes fair use under Section 107 of the US Copyright Act. The question is whether, between November 17, 2026, and December 30, 2027, inclusive, at least one US federal appellate court will issue a precedential ruling holding that training a generative AI model on copyrighted web content without authorization qualifies as fair use under US copyright law. Several significant cases provide context. In June 2025, the US District Court for the Northern District of California issued decisions in Bartz v. Anthropic and Kadrey v. Meta Platforms. In Bartz, the court found that using copyrighted books to train generative AI models could constitute fair use, while distinguishing the training process from the unlawful acquisition and retention of pirated copies. In Kadrey, the court granted Meta summary judgment on the training-related copyright claims based on the evidence presented, while emphasizing that AI training is not categorically protected by fair use. These decisions were issued by district courts rather than appellate courts. The Bartz litigation subsequently resulted in a settlement approved in July 2026. On September 29, 2026, the US Court of Appeals for the Third Circuit issued a significant appellate decision in Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. The court affirmed that ROSS Intelligence's use of copyrighted Westlaw editorial headnotes to train a competing, non-generative AI legal research system did not qualify as fair use. Importantly, the decision distinguished the non-generative system at issue from generative AI models and did not categorically determine whether training generative AI on copyrighted material constitutes fair use. Other copyright lawsuits involving generative AI developers, including litigation concerning OpenAI and Meta, continue to raise questions about the legality of training models on copyrighted internet content. These cases could produce further appellate decisions before the end of 2027. This forecast concerns whether a US federal appellate court will issue an affirmative fair-use ruling covering the unauthorized use of copyrighted web content to train generative AI models. It does not require a nationwide precedent, a Supreme Court decision, or a ruling establishing that all AI training is fair use. A qualifying holding may be limited to particular facts or circumstances, provided it expressly determines that the relevant training use constitutes fair use. Background sources include the US Copyright Office's summaries of Bartz v. Anthropic and Kadrey v. Meta, and the Third Circuit's September 2026 opinion in Thomson Reuters v. Ross Intelligence.

How it resolves

The question resolves YES if, between November 17, 2026, and December 30, 2027, inclusive, a US federal court of appeals issues a published, precedential majority or controlling opinion that expressly holds that using copyrighted content obtained from publicly accessible websites or other internet-accessible sources to train a generative artificial intelligence model without authorization from the copyright holder constitutes fair use under 17 U.S.C. § 107, under the factual circumstances considered by that court. A qualifying ruling must satisfy all of the following substantive conditions: It must be issued by one of the thirteen US federal courts of appeals, including the Court of Appeals for the Federal Circuit where applicable. It must address the use of copyrighted material in training a generative AI model, such as a large language model, image-generation model, audio-generation model, or another model capable of generating new expressive content. The relevant copyrighted material must have been obtained from the internet, including publicly accessible websites, online publications, internet-hosted repositories, or internet-accessible collections of copyrighted material. The court must determine that at least one challenged act of copying or using that content for model training, without the copyright holder's authorization, is protected by the fair-use doctrine. The determination must be part of the court's holding and necessary to its disposition of the relevant issue, rather than merely dicta, speculation, or a description of another court's reasoning. A ruling may qualify even if it is narrowly limited to a specific dataset, model, acquisition method, training practice, or set of copyrighted works. It may also qualify if the court determines that the training use itself is fair use while separately finding that acquiring or retaining unauthorized copies infringed copyright. The court need not find that every aspect of the defendant's conduct was lawful. The question resolves NO if no qualifying appellate decision is issued by December 30, 2027. It also resolves NO if appellate courts address AI-training copyright disputes but exclusively reject fair-use defenses, decline to reach the fair-use question, remand for further proceedings without making an affirmative fair-use holding, or issue only nonprecedential opinions that satisfy the substantive conditions. The primary resolution sources will be the official published opinions and orders of the US federal courts of appeals, as available through their official websites and the US Courts system. Court opinions reproduced by the Legal Information Institute at Cornell Law School, Google Scholar, or established legal reporting services may be used to locate or corroborate qualifying decisions. Where secondary reporting conflicts with the text of a judicial opinion, the official judicial opinion will control. The publication or filing date of the qualifying appellate opinion will determine whether the deadline is satisfied. The decision must be issued no later than December 30, 2027, using the date recorded by the issuing court. Subsequent petitions for rehearing, requests for Supreme Court review, or later reversals will not retroactively change a YES resolution, unless the qualifying opinion is withdrawn, vacated, or superseded before the resolution is finalized. If an opinion is withdrawn or vacated before final resolution, it will not count unless a replacement qualifying opinion is issued by the deadline. The question will not be annulled merely because relevant litigation settles, is dismissed, or does not reach an appellate court. In those circumstances, it will resolve NO unless another qualifying decision is issued. Annulment would be appropriate only if extraordinary circumstances make it impossible to establish the relevant judicial record reliably.

Fine print

For this question, 'US appeals court' means a federal court of appeals, not a state intermediate appellate court, a federal district court, or the US Supreme Court. A Supreme Court ruling alone therefore does not satisfy the question. 'Copyrighted web content' includes copyright-protected text, articles, books, images, photographs, music, and other expressive works made available through internet-accessible sources. The underlying works need not have originated online. However, the qualifying ruling must concern copies obtained through an internet-accessible source rather than exclusively through offline purchases or physical scanning. Content behind a subscription or paywall may qualify if accessed online. The fact that material was unlawfully uploaded to the internet does not automatically disqualify it, although the appellate court must affirmatively find the relevant training use to be fair use. 'Training' includes pretraining, continued pretraining, fine-tuning, and other processes that modify an AI model's parameters using copyrighted works. Merely retrieving copyrighted content at inference time, generating infringing outputs, or indexing content for conventional search does not qualify unless the court separately makes the necessary fair-use holding concerning generative AI model training. A decision concerning only a non-generative AI system, such as the system at issue in Thomson Reuters v. Ross Intelligence, does not qualify. Likewise, an appellate court's statement that generative AI training is transformative does not qualify unless the court also holds that the challenged training use satisfies the fair-use defense after considering the applicable statutory factors. An appellate decision affirming a district court's fair-use ruling qualifies if the appellate court expressly adopts that determination as part of its own holding. A decision affirming on unrelated grounds without reaching fair use does not qualify. A published, precedential opinion issued in an interlocutory appeal can qualify; the underlying litigation need not be fully concluded. For avoidance of doubt, 'before 31 December 2027' means no later than December 30, 2027, rather than December 31, 2027. The forecast starts on November 17, 2026, so earlier appellate rulings do not count.