Anthropic’s $1.5B copyright deal leaves AI fair use unresolved

A US federal judge on July 20, 2026, approved Anthropic’s $1.5 billion settlement of a copyright lawsuit brought by authors who accused the AI startup of using pirated books to train its Claude models. The agreement, first reached in 2025, is the largest known copyright settlement involving generative AI and the clearest indication yet of what copyright risk could cost companies building large language models.

The settlement does not answer the biggest legal question in the industry: whether training AI models on copyrighted material is considered as fair use. Instead, the settlement attempts to assign a value to this uncertainty. For developers, investors, and publishers, the case provides the first real means of valuing the repercussions of disputes related to training data.

This lawsuit, filed under the name Bartz et al. v. Anthropic PBC (Case No. 3:24-cv-05417) with the United States District Court for the Northern District of California, can have widespread implications for issues surrounding copyright and how such disputes are resolved, even though it might not establish binding legal precedent.

Fair use remains unsettled

The central point in the legal conflict is the concept of fair use. The US Copyright Office defines it as:

…a legal doctrine that permits the use of copyrighted material without the copyright owner’s permission under certain circumstances, with courts evaluating each case based on factors such as the purpose of the use, the nature of the copyrighted work, the amount used, and the effect on the market for the original work.

It is still uncertain whether that protection applies for training AI.

Peter Henderson, a Stanford JD/PhD student and co-writer of Foundation Models and Fair Use, claims the issue has not been sorted out yet.

There’s uncertainty about how lawsuits will come out in this area.

– Peter

Such uncertainty is not only theoretical. Researchers led by Henderson found that, through either simple prompt modifications or few derivations, GPT-4 can provide long segments of text from the book, “Oh, the Places You’ll Go!” written by Dr. Seuss and from the book “Harry Potter and Sorcerer’s Stone” even if one keeps some minimal modifications to the wording of the prompts.

This and other similar findings have been leveraged to make arguments that some AI models keep their protected expression rather than just learn patterns from their training material.

The US Copyright Office has reached a similar conclusion, stating that whether AI training can be classified as fair use or not depends on the unique facts surrounding each case rather than being governed by a general legal principle.

A benchmark for future AI lawsuits

The legal proceedings started in the year 2023, when some authors claimed that Anthropic had trained its systems with the use of illegal books. In 2025, the company decided to settle litigation for $1.5 billion following Judge William Alsup’s conclusion that AI training can indeed be treated as fair use, however, other claims related to acquiring pirated books still could continue. The final approval that made it the biggest copyright settlement in generative AI was provided by Judge Araceli Martinez-Olguin on July 20th, 2026.

The timing matters due to numerous notable lawsuits that are ongoing in US courts, such as the case initiated by The New York Times against OpenAI and Microsoft, multiple lawsuits initiated by writers against Meta, as well as the lawsuits initiated by different publishers, writers, and media organizations. Apart from those cases, Anthropic’s deal provides some immediate benchmark while leaving the issue of legality unsettled.

The market is watching

In terms of the AI industry, this settlement signifies that copyright risk is shifting from being an intangible legal concern into an actual cost of business. Organizations and enterprises are already spending billions on their chips, cloud systems and computing capabilities, and soon enough access to defensible training data will be yet another competitive edge.

Previously, Cryptopolitan published articles about Anthropic’s growing partnerships in computing, investment in AI research, and competition with Chinese foundation model developers. All those things indicate a bigger change: to succeed in AI, it is necessary not only to focus on computing and performance of the model, but also to have good data that is legally protected, in order to provide a successful outcome.

Perhaps the courts will one day decide where fair use ends and licensing starts. In the meantime, Anthropic will likely remain an unofficial marker of the industry price for uncertainty, as well as a reminder to many AI companies that settling is sometimes less risky than waiting for a ruling by the courts.