The impact of copyright settlements and antitrust fines on the AI evaluation ecosystem

Recent copyright settlements and antitrust fines mark potential turning points for how AI systems may be evaluated, governed, and trusted. From the perspective of AI evaluations, they signal that such evaluations can no longer focus exclusively on technical performance, or assessments of factuality, bias, and robustness, but must also account for data provenance, legal compliance, and market power.

Between September and October 2025, Anthropic agreed to pay approximately $1.5 billion to settle claims from authors and publishers that it had trained its large language model on pirate websites LibGen and Pirate Library Mirror (PiLiMi). A federal court had earlier held that training on legally purchased books might qualify as fair use, but using pirated libraries did not. As part of the settlement, Anthropic has agreed “to destroy all copies of pirated books in its possession” and pay rightsholders $3,000 per title. This case illustrates that the legality of AI training depends not just on the model’s purpose, but also on the lawful acquisition of that data. Other lawsuits against AI companies are also underway, including against OpenAI, similarly alleging copyright infringement.

During this time, the European Commission (EC) fined Google €2.95 billion (US $3.45 billion) for abusing dominance in its ad-tech business. The EC ordered the company to stop favoring its own ad exchange and warning that stronger remedies, including structural separation, could follow. While the Google case was not explicitly concerning AI, it reinforces that regulators, particularly those in Europe, are prepared to levy multi-billion-dollar penalties for anti-competitive or opaque practices in digital markets.

Together, these actions reshape the context for AI evaluation in four key ways:

1. Expanded scope of red teaming events: Red teaming events, like those that Humane Intelligence organizes, can now include data provenance-related exploits, such as testing whether models reproduce copyrighted material, reveal training sources, or expose data-leakage pathways. These developments enable security through legality as a new red teaming frontier. Likewise, algorithmic bias bounties can evaluate issues like copyrighted or unlicensed training data. 
2. Elevated data provenance and licensing as central to AI evaluations: Evaluators may now need to verify whether training datasets were lawfully obtained and properly licensed, before conducting an assessment. Benchmarking frameworks that once prioritized accuracy or fairness may need to include dataset-level compliance checks and metadata documentation, tracking where data came from, who owns it, and under what terms it was used. Clean data pipelines could become a competitive differentiator, while unverified data sources become a liability.
3. Expanded conception of responsible AI evaluation: Legal and ethical risk may become explicit dimensions of model assessment. Evaluators may test whether models reproduce copyrighted material, use unlicensed data, or create unfair competitive advantages. Organizations adopting external models may likely demand warranties and indemnities that all training data and market behavior comply with law. As such, the AI evaluation environment is increasingly incorporating legal components, integrating legal due diligence alongside model-card transparency and safety testing.
4. Increased costs and changed incentives: Licensing data rather than scraping it will likely increase costs for AI development, raising barriers for small firms but potentially improving dataset quality and accountability. The Anthropic case encourages the growth of licensed-data marketplaces, while the Google ruling pressures dominant firms to separate or disclose more about their data and algorithmic systems. Evaluation teams may therefore confront fewer, but more documented, models to benchmark, as companies invest in transparent governance to pre-empt regulatory action.

For evaluators and policymakers, the message suggests that legal risk, data rights, and competition dynamics are increasingly critical to measuring an AI system’s integrity. These legal shifts raise demand for third-party, rights-respecting evaluators. 

At the same time, the debate on copyright infringement remains a turbulent one, with competing developments. A U.K. court recently ruled in favor of Stability AI in a copyright case brought by Getty Images, offering competing precedent in the debate over AI training data. Getty had accused Stability of copying millions of its photos to train Stable Diffusion and of reproducing its watermark in generated images. The court dismissed all copyright claims, finding no evidence of direct copying, though it acknowledged some limited trademark issues. The decision is considered a boost for AI developers, suggesting that models trained on copyrighted material may avoid infringement if they do not reproduce works verbatim.

Ultimately, in the years ahead, trust in AI may depend as much on how models are constructed and governed, as on what they can do and output. Evaluation frameworks may evolve to include governance and compliance metrics, verifying lawful data sources, auditing licensing statements, and tracking vendor accountability. At the same time, the discussion remains a volatile one, and future legal decisions may shift the trajectory of accountability and correspondingly, AI evaluations.