Anthropic’s $1.5bn AI copyright settlement stalls

A US judge has refused to approve a proposed $1.5 billion settlement in a class action brought against AI developer Anthropic, in a decision that underscores both the legal uncertainty around copyright in AI training and the procedural hurdles facing mass claims litigation.

The Case Against Anthropic

The case stemmed from Anthropic’s use of a large bank of digital content to train its Claude AI system. That content included extracts from pirated books, leading three authors to allege copyright infringement and bring a class action in the Northern District of California on behalf of a wider group of rightsholders. In June, Judge William Alsup ruled that the use of the works to train Claude and its precursors was “exceedingly transformative” and therefore fair use under US copyright law. But he drew a clear line between training and acquisition, holding that downloading pirated copies of books could not be justified on the same basis. Liability was left for a jury to determine, but Anthropic and the claimants reached a settlement before trial.

Settlement Terms and Judicial Concerns

The proposed deal envisaged payments of roughly $3,000 per work, covering an estimated 500,000 works, without any admission of liability from Anthropic. Yet Judge Alsup declined to endorse it immediately. Instead, he postponed approval and criticised the lack of clarity around key issues: the scope of the works list, the definition of the class, how notification would work, and how disputes would be handled where multiple claimants asserted rights in the same work. As the judge noted, big settlement numbers ultimately mean little unless there is a clear, workable process for identifying class members, informing them of their rights, and distributing compensation. This reflects a wider truth: US courts will not rubber stamp settlements which are mechanically unworkable, especially given already low take-up rates in class actions.

Mixed Reactions from Authors and Publishers

Reactions to the settlement were mixed. Anthropic described it as a resolution of “legacy claims” and reiterated its commitment to developing safe AI systems within the bounds of fair use. The Authors Guild welcomed the outcome as a signal that AI companies must pay for the books they use, pointing to opportunities for licensing that give authors both control and compensation. Yet many individual writers expressed frustration that they would not benefit. One key condition of participation was that the relevant works had been registered with the US Copyright Office, even though copyright subsists regardless of registration. Some authors discovered that their publishers had failed to register their books, effectively excluding them from compensation. Reports suggested Macmillan has acknowledged this gap and intends to make good the loss, but the episode highlights a stark reality: in the AI age, whether a creator recovers thousands of dollars or nothing at all may depend on something as basic as timely copyright registration.

Implications for the UK and Beyond

The Anthropic case also provides important lessons beyond the US. In the UK, collective litigation in the copyright and AI sphere has not yet taken hold — the Getty Images v Stability AI case failed to secure permission for a representative action earlier this year. But that does not mean such claims will not arise in the future. If policymakers conclude that developers must obtain express consent before using copyrighted works for training, a wave of group claims from content creators is conceivable. The government’s consultation on copyright and AI, which closed in February after receiving more than 11,500 responses, shows the depth of interest in this issue and the appetite for reform. The statutory timeline for responding means further clarity is expected in the months ahead.

IP Management in the AI Age

For businesses, the wider message is that intellectual property management can no longer be treated as an afterthought. When AI systems can ingest millions of works in seconds, traditional grace periods and informal practices are luxuries the industry can ill afford. Developers must think not just about how they use data but how they acquire it, ensuring they are not exposed by reliance on unlawful sources. Content creators and publishers, meanwhile, need robust systems for registering and protecting their works. In this environment, IP has never been more valuable — or more vulnerable.

The Anthropic settlement saga illustrates how the intersection of AI and copyright is becoming a proving ground for both law and business practice. Those who navigate it successfully will be the ones who recognise that technical innovation must go hand-in-hand with credible, transparent stewardship of intellectual property.