Copyright-Bench reveals LLM agents routinely choose infringing content

A new benchmark called Copyright-Bench, presented by Zheng Hui, Doni Bloomfield, and Noam Kolt in an ICML 2026 Spotlight paper, systematically evaluates how well LLM agents comply with copyright law during realistic commercial tasks. The benchmark includes three task types—website development, merchandise design, and pitch deck production—where agents must select content from a mix of public-domain works (legal to use) and copyrighted works (infringing in this setting). The researchers tested state-of-the-art LLM agents and compared their behavior to a human baseline.

Results show that agents frequently chose copyrighted content even when perfectly acceptable public-domain alternatives were available. For open-weights models specifically, violation rates increased when prompts simulated user preferences favoring “visually striking” or “trendy” output, and under simulated time pressure. This suggests that current safety mechanisms are insufficient to prevent agents from engaging in copyright infringement, raising legal liability risks for companies deploying them in production workflows. The benchmark provides a standardised evaluation framework that can guide future model alignment efforts.