Claude's new system prompt doesn't want to reproduce song lyrics (simonwillison.net)

Anthropic has announced significant updates to Claude's system prompts, particularly focusing on copyright compliance and user interaction guidelines. The most noteworthy addition is a strict policy against reproducing song lyrics, poems, or passages from copyrighted texts, regardless of context. This change arises in light of recent legal challenges from music publishers Sony Music and Warner Chappell, which highlight ongoing concerns around copyright infringement in AI model training. Claude will now automatically decline requests for such content in a conversational setting, offering alternatives like analysis or description instead. Additionally, improvements have been made to Claude's interaction style and abusive behavior handling. The model will now respond with concise, direct answers without unnecessary qualifiers, and it will no longer apologize excessively or end conversations if a user is rude, aiming instead for accountability without submissiveness. This reflects a broader shift in design philosophy within Claude's architecture, emphasizing respectful engagement while refusing to create or modify recognizable copyrighted material, expanding its scope to visual works and further refining its operational parameters. These adjustments underscore the evolving landscape of ethical AI development and the growing importance of robust copyright management in AI-generated content.