Anthropic Updates Claude System Prompts to Strictly Block Copyrighted Lyrics and Text
Simon Willison analyzed Anthropic's updated system prompts for Claude models, revealing strict new guardrails prohibiting the AI from reproducing copyrighted song lyrics, poems, or book passages. The updated system prompt explicitly instructs Claude to decline these requests even when users paste text line-by-line or claim it is their own work. This change demonstrates how AI labs are actively strengthening system-level prompt engineering to avoid copyright infringement risks and legal challenges from music publishers and authors. It also highlights Anthropic's unique transparency in publishing and maintaining versioned histories of its official system prompts. The guardrails explicitly allow works first published before 1929, but instruct Claude to rely on its own internal knowledge of publication dates rather than user statements. Additionally, once Claude refuses a copyright-restricted request in a session, it is programmed to keep rejecting narrower or reworded attempts throughout that entire conversation.
## BACKGROUND
In large language models (LLMs), a system prompt acts as a foundational set of top-level instructions that define the AI's role, behavioral boundaries, and safety policies before user interactions begin. AI developers face increasing legal pressure and copyright lawsuits over the unauthorized reproduction of protected literary and musical works. Anthropic publicly hosts its consumer system prompts on its documentation site, allowing developers to inspect prompt modifications over time.