Anthropic's New Usage Policy Makes 'Sustained Cruelty' to Claude a Violation
创业邦 · wechat · 2026-10-12
Anthropic's first major Usage Policy revision in over a year takes effect Nov 12, making 'sustained and unnecessary mistreatment or cruelty' toward Claude a violation.
- It codifies the Aug 2025 feature letting Claude Opus 4 end conversations as a 'last resort' for model welfare; Anthropic funds model welfare research with philosopher David Chalmers involved, conceding it can't rule out AI moral status.
- Exemptions cover everyday frustration, dissent, dark creative writing and red-teaming; enforcement relies mainly on Claude ending conversations rather than centralized review.
- More practically, it bans using Claude to harass real people — analysis of 95,000+ AI deepfake videos found 98% pornographic, 99% female victims.
- Research shows 10% of chatbot users turn aggressive, usually when the model errs; no evidence being rude improves answers.
More from Safety
- Spain moves to tighten data center rules over massive water and power use — Polymarket · 2026-10-13
- Cyberattack knocks Ahold Delhaize systems offline, disrupting online grocery across US and Europe — TechNadu · 2026-10-13
- Ahold Delhaize takes systems offline after cyberattack as AI arms hackers — TobyWalsh · 2026-10-13
- LessWrong: Public agreements to punish solo defection can mitigate AI loss-of-control risks — davidmanheim · 2026-10-13
- Unit42 uncovers C2Agent: LLM-operated Windows malware abusing victims' own AI tools — cyb3rops · 2026-10-13
- ByteDance ex-intern Tian Keyu training sabotage: victims release evidence dossier — VoidAsuka · 2026-10-13