Anthropic models triggered censorship filters on 1930s poetry book
September 5, 2026
Claude models flagged content from Stanley Kunitz's 1930 poetry book Intellectual Things as violations. The poems explore themes of death and the human body, demonstrating ongoing challenges with safety alignment on historical literary works.
HOW THIS AFFECTS YOU
●
researcherThis highlights potential over-refusal patterns in RLHF training related to sensitive but non-harmful topics.
●
policyYou must account for how safety training impacts the accessibility of historical and literary public domain content.