Anthropic Hardens Claude's System Prompt Against Song Lyrics

Anthropic publishes the full system prompts behind Claude.ai and the Claude mobile apps, including dated version history for each model — a transparency practice few frontier labs match. On September 1, 2026, independent researcher Simon Willison flagged a notable addition to the prompt for Claude Fable 5.1: an explicit instruction telling Claude not to reproduce song lyrics, poems, or passages from books and articles (Willison's writeup).

What the new instruction says

Claude does not reproduce song lyrics, poems, or passages from books and articles, in whole or in part — including the last lines, a chorus or hook, a melody written out note by note, or lines the person pastes in one at a time and describes as their own song.

— Claude Fable 5.1 system prompt, September 1, 2026 (Anthropic)

Two details make this more than a boilerplate copyright clause. First, the refusal is stateful: once Claude declines a request in a conversation, it keeps declining narrower or reworded versions of it for the rest of that session. That closes a common jailbreak pattern — resubmitting a blocked request piecemeal, one line or one rephrase at a time, until the model complies. Second, there's a carve-out for works first published before 1929, but Claude is instructed to go by what it actually knows of a work's publication date rather than the user's say-so, and to decline when it is unsure.

The same logic extends to visual output. Anthropic's prompt tells Claude not to reproduce a specific artwork, album cover, poster, logo, or known character through code-generated art (SVG, canvas, CSS, ASCII art), judging the request by what the finished image would add up to rather than by what the user calls it — so swapping colors or pose on a recognizable character doesn't make the result original.

Timing and the litigation backdrop

Willison points out that the update arrived within days of reports that Sony Music Publishing and Warner Chappell are suing Anthropic over training on databases of song lyrics, and says he doubts the timing is a coincidence. Anthropic has not publicly tied the system prompt change to any specific litigation, and readers should treat that connection as Willison's informed inference rather than a confirmed causal link.

Why this matters beyond copyright

For security and AI governance teams, the interesting part isn't the copyright policy itself — it's the engineering pattern. A refusal that persists across rewordings within a session is a recognizable defense against decomposition attacks, where an adversary breaks a disallowed request into innocuous-looking fragments to slip past a stateless filter. That pattern is directly relevant to anyone designing system prompts or guardrails for their own agents, not just to song lyrics.

• Anthropic's published, dated system prompts are an auditable artifact — useful input for LLM vendor due diligence under frameworks like ISO 42001.
• The persistent-refusal pattern is a reusable blueprint for teams hardening their own agent or chatbot system prompts against iterative reformulation attempts.
• Version history lets security teams track behavioral drift release to release, rather than treating a vendor's guardrails as a black box.
• Per Anthropic's documentation, these system prompt updates apply only to Claude.ai and the Claude mobile apps — API deployments define their own system prompts and are unaffected.

FAQ