Anthropic bans Claude from reproducing song lyrics

Summary

Anthropic added an aggressive anti-evasion clause to its updated Claude system prompt (Fable 5.1). This clause prohibits the reproduction of song lyrics, poetry, and copyrighted visual characters. This content will not be generated even via SVG, ASCII art, or rephrased prompts. This change came days after Sony Music Publishing and Warner Chappell filed a training data lawsuit against the company.

Why it matters

The timing makes it difficult to ignore the legal motivation. A prompt-level patch is a cheap and fast way to signal compliance to courts and juries without altering the underlying model weights. Additionally, this reveals that the prompts published by Anthropic are only a slice of the actual instruction stack. Critical behaviors like end_conversation reside in unpublished feature-based blocks. This complicates any claims of transparency.

Highlights

• The new song lyric ban explicitly covers 'lines pasted individually by a person and defined as their own song.' This closes a common jailbreak vector.
• A cute example quoted from the system prompt (Sonic → axolotl doing a skateboard trick) could subtly steer the model toward this specific working method in future conversations.
• Anthropic's documentation site is LLM-friendly (adding .md extensions for Markdown). This makes automating prompt diffs extremely easy. However, the full runtime prompt remains partially hidden.

Critical take

Simon Willison rightly points out that the prompt-level fix is reactive and cosmetic. This approach tells the model to refuse the request. However, it does not address whether the model was trained on these lyrics. Furthermore, core behavior rules like end_conversation, memory, and web search remain in unpublished layers. Therefore, publishing the core prompt as 'transparency' carries more of a marketing nature than accountability.