The Anthropic authors' settlement: what the court approved, and what it did not decide

What happened

Authors brought a class action in the United States alleging that Anthropic infringed copyright in connection with building its large language models. As reported by the Associated Press, the parties proposed a settlement of approximately $1.5 billion covering a large set of book titles, and the Los Angeles Times reported that a judge granted final approval in July 2026.

The distinction that matters most is between two different acts. One is using lawfully obtained copyrighted works to train a model, which the court treated as capable of qualifying as transformative fair use under US law. The other is acquiring and retaining copies of books from pirate sources to build an internal library, which was treated as a separate and legally weaker position. The settlement resolved claims arising from the second category rather than establishing that all AI training is lawful.

A settlement is a negotiated resolution. Approval means a court found the terms fair, reasonable and adequate for the class; it is not a ruling that every allegation was proven, and it does not create binding precedent in the way a fully litigated appellate judgment would. Other AI copyright cases involving different defendants, different works and different acquisition practices remain unresolved.

Why it matters

The scale of the payment is significant because it puts a concrete price on data sourcing practices. For developers, it shifts the question from whether training data matters to how provenance is documented, licensed and retained.

For authors and other rightsholders, it demonstrates that collective action can produce compensation at scale, and that the weakest link in a developer's position may be procurement rather than the modelling itself.

For readers and consumers, the case is part of a broader shift towards expecting disclosure about what goes into a model. Transparency about training data is becoming a compliance obligation in some jurisdictions rather than a voluntary gesture.

What consumers should watch for

Resist sweeping claims in either direction. This outcome neither authorises unrestricted scraping nor establishes that training on copyrighted works is unlawful in general. The specific facts about how material was obtained were central.

If you are an author, check whether any settlement or claims process applies to your titles, what the deadline is, and what rights you release by claiming. Verify claims administration details through official channels; large settlements attract impersonation attempts.

Look for provenance signals in the tools you use: published documentation of training sources, licensing arrangements with publishers, and opt-out mechanisms for rightsholders.

European context

European law approaches this differently and does not have a general fair use doctrine. Instead, the Directive on Copyright in the Digital Single Market provides specific exceptions for text and data mining. One applies to research organisations and cultural heritage institutions for scientific research. A broader exception permits text and data mining on lawfully accessible content, but rightsholders may expressly reserve their rights, for example through machine-readable signals. Lawful access is a precondition in both cases, which is why obtaining material from pirate sources sits outside the exception regardless of the purpose.

The EU Artificial Intelligence Act adds obligations for providers of general-purpose AI models, including a policy to respect copyright and rights reservations, and publication of a sufficiently detailed summary of training content. These are regulatory duties owed to authorities and the public rather than a direct compensation mechanism for individual authors.

For redress, European rightsholders generally rely on national copyright enforcement, collective management organisations, and licensing negotiations. Consumer collective redress under the Representative Actions Directive is directed at consumer law breaches and is not the usual route for authors' copyright claims. None of this is a comment on any specific matter or party.

Practical next steps

Authors and publishers should record their catalogue clearly, including ISBNs, publication dates and territory of rights, so that eligibility for any claims process can be established quickly.

Consider setting an explicit rights reservation for text and data mining on your own websites and with your distributors, in a machine-readable form where possible, and keep evidence of when it was applied.

If you find your work reproduced without permission, preserve evidence with dates and, where relevant, use the notice mechanisms of the platform or service concerned.

For organisations building or deploying AI systems, document the provenance and licence basis of training and fine-tuning data, and retain that documentation. Obligations and available remedies differ by country, so take local advice before relying on any single framework.