AI and Copyright Infringement

Every major AI lab that trained a large language model on scraped or scanned text is now, or has recently been, a defendant in a copyright suit over it. As of 2026-07-12, that body of litigation has produced several first-instance rulings, at least one enormous proposed settlement, and zero final, appeal-exhausted answers to the central question: is training a generative model on copyrighted text fair use? District courts have split depending on how the training data was acquired and what kind of product it was used to build. Nothing below should be read as a settled, industry-wide rule.

1. 1. Purpose and character Is the use transformative, meaning it does something meaningfully different with the work, or commercial and duplicative of the original purpose?
2. 2. Nature of the work Is the copyrighted work factual or highly creative? Creative works get more protection.
3. 3. Amount used How much of the work was copied, and was that amount reasonable for the purpose?
4. 4. Market effect Does the use substitute for the original in the market, or otherwise damage its value?

Courts apply all four factors together; no single factor controls, and different judges have weighed them differently in the cases below.

The suits without a resolution yet

Authors Guild v. OpenAI, Inc. is active and in discovery in the Southern District of New York, with no final judgment. In October 2025, the judge denied OpenAI summary judgment on a narrower theory, that ChatGPT's plot summaries and outputs of the plaintiffs' novels could themselves be infringing, separate from the training-data question. That ruling addressed only output, not whether training on the books was fair use, which remains pending.

The New York Times Co. v. Microsoft Corp. and OpenAI is likewise active, in discovery and pre-summary-judgment posture with no trial date set. An April 2025 order resolved motions to dismiss and let the main claims proceed. Since then, the parties have fought over preserving and producing ChatGPT user logs, the Times has filed a 2026 sanctions motion alleging OpenAI concealed search capabilities over its training data, and in June 2026 the Times moved to amend its complaint to add allegations against Microsoft. Summary judgment briefing was reportedly wrapping up around April 2026, but as of 2026-07-12 no summary judgment ruling had issued in the sources checked. This case is moving quickly and its status should be reconfirmed before relying on it.

Where courts have actually ruled, and how far those rulings reach

Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. is the one outright rejection of a fair-use defense in this group. On February 11, 2025, the court granted Thomson Reuters partial summary judgment, finding Ross's use of Westlaw headnotes to train a competing legal-search tool non-transformative, commercial, and directly market-substituting. That ruling is on appeal to the Third Circuit as of 2026-07-12 and involved a non-generative search product, a materially different fact pattern from a general-purpose language model trained on books or news text.

Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, Inc. both went the other way on the training question itself. In June 2025, one judge held that training Claude on lawfully acquired, scanned books was "exceedingly transformative" fair use, but that Anthropic's downloading of over seven million pirated books to build a permanent library was not fair use and created separate liability. A proposed $1.5 billion class settlement followed, reported as the largest copyright class settlement on record; a fairness hearing in May 2026 did not result in final approval, and none had issued as of the most recent sources checked. Around the same time, a different judge granted Meta partial summary judgment in Kadrey, holding that training Llama on the plaintiffs' books was "highly transformative" fair use, while noting in dicta that a market-dilution theory the plaintiffs hadn't adequately developed might have changed the result. As of June 2026, the Kadrey plaintiffs had moved to certify that ruling for interlocutory appeal to the Ninth Circuit, partly to resolve tension with how the Bartz court treated pirated source copies. No appellate ruling exists yet on either case.

Read together, these rulings suggest transformative training on lawfully obtained text has fared well so far, while how the data was acquired, and what kind of product it powers, keeps producing different outcomes. None of that is a final rule, and readers should treat the fair-use question for AI training as genuinely open until an appellate court decides it.