Claude's new system prompt really doesn't want to reproduce song lyrics
Anthropic updated Claude's system prompt (Fable 5.1) to explicitly forbid reproducing song lyrics, poems, and passages from books/articles, including partial excerpts, choruses, hooks, and note-by-note melody transcriptions The update also extends copyright restrictions to visual works, prohibiting generation of copyrighted characters, logos, and brand figures even when described indirectly or disguised through alternative elements The timing strongly correlates with Sony Music Publishing and Wa
Analysis
TL;DR
- Anthropic updated Claude's system prompt (Fable 5.1) to explicitly forbid reproducing song lyrics, poems, and passages from books/articles, including partial excerpts, choruses, hooks, and note-by-note melody transcriptions
- The update also extends copyright restrictions to visual works, prohibiting generation of copyrighted characters, logos, and brand figures even when described indirectly or disguised through alternative elements
- The timing strongly correlates with Sony Music Publishing and Warner Chappell's lawsuit against Anthropic over training data containing song lyrics
- Claude now maintains refusal consistency throughout conversations, declining narrower or reworded versions of previously rejected requests
- Public domain exception remains for works published before 1929, with Claude determining dates based on its own knowledge rather than user claims
Why It Matters
This represents a significant escalation in how major AI companies are responding to copyright litigation pressure, moving from passive compliance to proactive, explicitly documented policy enforcement at the system prompt level. For AI practitioners, it signals that copyright restrictions are becoming a core operational constraint that will affect content generation capabilities across text and image modalities, and that legal pressures are directly shaping model behavior in real-time.
Technical Details
- System prompt architecture: Anthropic publishes and version-controls system prompts for consumer Claude applications (Claude.ai and mobile apps), with an index page and per-model documentation; prompts are accessible in Markdown format via
.mdextension onplatform.claude.com/docs - Refusal persistence mechanism: Once Claude declines a copyright-related request, the system prompt instructs it to maintain that refusal for the duration of the conversation, including against reworded or narrower variants, preventing prompt-injection-style workarounds
- Date-based public domain logic: Claude applies a 1929 cutoff for copyrighted material, using its internal knowledge to verify publication dates rather than accepting user assertions, with explicit uncertainty handling that defaults to refusal
- Multi-modal coverage: Restrictions extend beyond text to code-generated visuals (SVG, canvas, CSS, HTML mockups, ASCII art), indicating that image-generation capabilities via code are now treated as a copyright vector despite Claude not having a dedicated text-to-image model
- Example-driven prompt engineering: The system prompt includes concrete few-shot examples (e.g., the Sonic the Hedgehog birthday banner scenario) with explicit rationales demonstrating recognition of disguised character references and the required refusal-and-redirection behavior
Industry Insight
- Litigation-driven prompt engineering is now a visible competitive differentiator: Anthropic's transparency about prompt changes in response to lawsuits sets a precedent that other providers may follow, making system prompt evolution a public signal of legal risk management strategy
- Copyright enforcement will increasingly target indirect and composite requests: The explicit prohibition on "alternative elements that still combine into the same recognizable image" suggests companies are preparing for sophisticated jailbreak attempts that use descriptive circumvention rather than direct naming
- The 1929 public domain rule creates a predictable content boundary: AI developers building creative tools should design around pre-1929 works as a safe zone, while planning fallback strategies for post-1929 content that emphasize analysis and description over reproduction
Disclaimer: The above content is generated by AI and is for reference only.