Anthropic’s $1.5bn settlement stalls as authors demand higher payouts

A federal court has delayed approval of Anthropic’s $1.5 billion copyright settlement after multiple authors challenged the proposed compensation structure, arguing the payout formula undervalues their intellectual property used to train Claude AI models.

The settlement, which would resolve class-action claims from thousands of authors whose works were allegedly used without permission in training data, faces objections from writers who contend the distribution mechanism favours quantity over quality and fails to account for the commercial value their specific works provide to Anthropic’s large language models.

According to court filings reviewed by Ars Technica, objecting authors argue that the settlement’s per-work payment structure—estimated at approximately $150 to $300 per book depending on final participation rates—bears no relationship to the actual value their creative works contribute to Claude’s capabilities. Several prominent authors have filed formal objections, claiming their bestselling titles generate disproportionate training value compared to obscure works that would receive identical compensation.

The dispute centres on fundamental questions about how to value copyrighted material in AI training datasets. Anthropic’s proposed settlement treats all written works as functionally equivalent, distributing funds based primarily on word count and publication status rather than market success, critical acclaim, or demonstrable impact on model performance.

Legal experts note the case could establish precedent for dozens of pending copyright actions against AI companies. OpenAI, Meta, and Stability AI all face similar litigation from authors, visual artists, and news organisations over training data practices. How courts value individual copyrighted works within massive training datasets remains legally unsettled territory.

The business implications extend beyond Anthropic’s balance sheet. If courts accept author arguments for differentiated compensation based on commercial value, AI companies may face substantially higher settlement costs in pending cases. This could fundamentally alter the economics of training frontier models, potentially requiring companies to negotiate individual licencing agreements for high-value content rather than relying on fair use defences or blanket settlements.

For publishers, the outcome determines whether they can extract premium rates for catalogue content used in AI training. Major publishing houses have largely stayed neutral in the Anthropic case whilst pursuing their own negotiations with AI companies, but author objections could strengthen their bargaining position by establishing that not all training data carries equal value.

Smaller AI companies face particular exposure. Whilst Anthropic’s $1.5 billion settlement represents roughly 15% of its total funding raised, similar proportional settlements could prove existential for startups without comparable financial backing. The case may accelerate consolidation as only well-capitalised firms can afford the legal and financial costs of resolving training data disputes.

The court has scheduled a fairness hearing for next month, where objecting authors will present arguments for rejecting or substantially modifying the settlement terms. Legal observers expect the judge to request additional economic analysis demonstrating how the proposed compensation correlates to actual value provided to Anthropic’s models.

Industry attention now focuses on whether Anthropic will revise the settlement structure to address author concerns or proceed to trial if objections prevent approval. The company has not publicly commented on the objections, though sources familiar with the matter suggest legal teams are exploring modified distribution formulas that might satisfy both parties whilst keeping total costs within the $1.5 billion framework.

The delay creates uncertainty for thousands of class members awaiting payment and leaves unresolved the broader question of how AI companies should compensate creators whose works fuel model development. Whatever framework emerges from this case will likely shape training data economics across the industry for years to come.