Key points

• Courts have started to separate two questions that were once run together: whether training a model on copyrighted works is a fair use, and whether acquiring and keeping those works in the first place was lawful.
• Early US district court rulings in 2025 favored developers on the training question where the works were lawfully obtained, but the reasoning was fact-specific, and the decisions are not binding on other courts.
• Allegations that a model reproduces or memorizes protected works in its outputs are proving a separate and, for some defendants, more difficult line of attack than the training claim itself.
• Outside the United States, courts in the UK and Germany have reached different conclusions on related questions, which means a global product faces a patchwork of exposure.
• Companies that license, fine-tune or deploy third-party models should focus on contractual protection, provenance records and output controls rather than waiting for a single definitive ruling.

Where the litigation stands

More than fifty copyright suits against AI developers are pending in US courts, brought by authors, news organizations, music publishers, visual artists and software developers. They have moved from pleading disputes to summary judgment and, in a few cases, to settlement. A year ago, commentators described the law as entirely unsettled. It is still unsettled, but the shape of the argument is clearer, and the questions that matter most are more precise.

The central doctrine is fair use under section 107 of the Copyright Act. Courts weigh four factors: the purpose and character of the use, the nature of the work, the amount used, and the effect on the market for the original. In AI cases the first and fourth factors do most of the work. Defendants argue that training is transformative because the model learns statistical patterns rather than republishing expression. Plaintiffs argue that training copies entire works without permission in order to build a product that competes with them.

What the 2025 district court decisions did and did not decide

In 2025, two federal judges in the Northern District of California issued summary judgment rulings in cases brought by authors. Both found that training a large language model on books was highly transformative. One court treated lawfully purchased books differently from a central library assembled from pirated copies, and allowed the claim about the pirated library to proceed toward trial. The other ruled for the developer largely because the plaintiffs had not built a record showing market harm, while signaling that a better record on market dilution could change the outcome. A separate decision in the District of Delaware, involving a non-generative legal research tool, rejected a fair use defense on its facts.

Read together, these rulings do not establish a rule that training is fair use. They establish that the answer depends on how the data was obtained, what the model does with it, and what evidence exists about market effects. They are also district court rulings. Appeals are pending in some matters, and other judges are free to reach different results. A settlement reported in one of the author cases, widely described as the largest publicly reported copyright recovery to date, also shows that acquisition risk can be priced and resolved even where the training question remains open.

Outputs, memorization and market substitution

The next wave of claims looks beyond training to what models produce. Plaintiffs increasingly allege that a model can regurgitate substantial portions of a protected work, or generate summaries and derivative content that substitutes for the original. News publishers and music rights holders have been particularly active here. These claims are attractive because they are easier to demonstrate with examples and because they connect directly to the fourth fair use factor. A model that reproduces a work on request looks less like learning and more like distribution.

Retrieval-augmented systems raise a related question. When a product fetches current articles and summarizes them at query time, the training defense is not available, since the use occurs at inference. Defendants instead rely on arguments about transformation, the limited nature of the excerpts shown, and the absence of a substitute for the original. Courts have not yet resolved these arguments.

The international picture

The United Kingdom's text and data mining exception is narrow and covers only non-commercial research, and the government has continued to consult on reform. A 2025 High Court decision in a case concerning an image generator largely rejected claims that the model itself was an infringing copy, while leaving open other questions about training that took place outside the UK. A German regional court has found that memorization of song lyrics in a model's parameters can amount to reproduction. In the European Union, the AI Act requires providers of general-purpose models to adopt a copyright compliance policy and to respect machine-readable rights reservations under the text and data mining exception in the Copyright Directive. A company with a global product should assume that the most restrictive plausible reading in any market where it operates will shape its risk.

The US Copyright Office's position

The Copyright Office published a pre-publication report in 2025 examining generative AI training. It concluded that some training uses will be fair and others will not, that commercial use of pirated or unlawfully accessed material weighs against fair use, and that market dilution is a legitimate consideration. The report is persuasive rather than binding, but courts and litigants cite it, and it reflects the questions regulators will keep asking.

What companies should do now

For developers, the immediate priority is a provenance record. Know what data went into each model, how it was acquired, what licenses or terms attached to it, and whether rights holders have opted out. Where datasets include material of uncertain origin, consider remediation or licensing before disputes arise. Output controls, including filters against verbatim reproduction and policies on prompts that request specific protected works, are now a standard part of a defensible position.

For companies that buy or fine-tune third-party models, the contract is the principal protection. Seek an intellectual property indemnity that covers both training claims and output claims, check that carve-outs for fine-tuned or modified versions are workable, and understand what the vendor will do if a court orders a model withdrawn or retrained. Our Kenji Watanabe-Price and Daniel Achterberg have written separately on the specific terms to negotiate, and our Technology Transactions practice regularly advises on them.

For rights holders, the developments suggest that licensing markets are forming faster than doctrine. Evidence of lost licensing revenue is becoming central to the fourth factor, so building a record of licensing offers and market transactions matters.

Litigation in this area will continue for years. The prudent course is to manage exposure through documentation and contracts now, while following the appellate decisions that will begin to settle the core questions. Corvane Ryle's Intellectual Property Litigation practice and Artificial Intelligence practice advise AI developers, deployers and rights holders on training-data strategy, licensing and disputes, and our litigators, including Celeste Varga, act in this area.