Anthropic updates Claude system prompt to block reproduction of song lyrics

Anthropic added a new clause to the system prompt for Claude’s consumer applications that blocks the model from reproducing song lyrics, poems, or book passages in any form. The clause also covers visual outputs, prohibiting the generation of any recognizable copyrighted character, logo, or design. After a refusal, the model continues to decline reworded requests for the rest of the conversation and offers to describe or analyze the requested work instead.

Engineers who integrate Claude.ai or the Claude mobile apps must now expect lyric-related requests to be refused and design their applications to handle that outcome gracefully. This may require adding fallback logic that redirects users to a description or analysis feature, or that blocks the request upstream. The change does not affect Claude Cowork or Claude Code, so internal tools remain unchanged unless they also adopt the updated prompt. Teams should review their prompt pipelines to ensure they do not inadvertently trigger the lyric block through rephrasing.

The block applies only to works that the model judges to be under copyright based on its internal knowledge of publication date; works published before 1929 are permitted. If the model is uncertain about a work’s date, it will err on the side of refusal, which can lead to over-blocking of borderline cases. Consequently, the system will not stop working for permissible older works, but it may refuse legitimate requests for recent works that the model misidentifies.