Anthropic publishes two safety plans, the legally binding one is weaker
sjgadler · x · 2026-10-10
Safety researcher Garrison Lovely points out that California law (SB 53, effective January) lets AI companies voluntarily create safety plans they are legally compelled to stick to—yet Anthropic, the self-styled "safety" company, doesn't do this. Instead it publishes two versions of its safety plan, with the weaker one being the legally binding version.
AI policy researcher Nathan Calvin adds that calling Anthropic's RSP "commitments" is misleading: if they wanted real commitments, they could put more content into their legally binding frontier safety framework under state law. As a transparency document about Anthropic's beliefs on best practices, the RSP is useful; as hard commitments giving policymakers confidence, it falls short.
Related event: Critics Say Anthropic's Safety Commitments Have Legal Loopholes(3 posts)→
More from Companies & People
- Founder: Letting Day One Ventures onto my cap table was my worst mistake — ctjlewis · 2026-10-12
- Anthropic and Resolution Abolished Take-Home Coding Tests as AI Makes Them Cheatable — BlackHC · 2026-10-12
- UC Irvine opens AI/ML faculty search with salaries up to $241,400 — StephanMandt · 2026-10-12
- Ex-OpenAI researcher Diogo Almeida left at ChatGPT's peak, now building TypeSafe AI — hardimanjames · 2026-10-12
- ThursdAI weekly: OpenAI drops 722 math papers in one night, Haiku 5.5 hits 10 cents — thursdai_pod · 2026-10-12
- Sarah Hooker cheers killing Leetcode interviews: low correlation with real engineering — 3scorciav · 2026-10-12