Protecting the wellbeing of our users

Source: Anthropic Date: 2025-12-18 URL: https://www.anthropic.com/news/protecting-well-being-of-users

Summary

Anthropic published a user wellbeing framework covering crisis response, sycophancy, and mental health safeguards. Suicide/self-harm: classifier directing to crisis resources (ThroughLine, 170+ countries); Opus 4.5 responds appropriately 98.6% of the time to high-risk requests. Sycophancy addressed: latest models “substantially outperform” previous versions and competitors on the open-source Petri evaluation. Partners: International Association for Suicide Prevention, Family Online Safety Institute. Claude.ai requires users to be 18+.

Implications

• Consumer trust / safety posture thread. Publishing specific performance metrics on mental health safeguards (98.6% appropriate response rate) is a rare quantitative claim about safety behavior. It’s also the kind of metric that mental health advocacy groups and regulators will hold Anthropic to.
• ThroughLine at 170+ countries. The global crisis resource integration is operationally significant — it means Claude’s crisis response is localizable, not just English/US-centric. 170 country coverage is broader than most crisis response infrastructure.
• Sycophancy as a named safety problem. Anthropic framing sycophancy as a user harm (telling users what they want rather than truth) and publishing Petri evaluation comparisons elevates it from an alignment quirk to a consumer protection issue. This is proactive — before regulators define it as a problem.
• 18+ age requirement. Claude.ai’s 18+ age gate is stricter than some competitors and relevant to the academic integrity and mental health safety discussions. It limits consumer market TAM but reduces regulatory risk from minors’ use.
• Watch: whether the 98.6% crisis response rate holds in adversarial testing; how the Petri sycophancy benchmark gets adopted by other labs; whether the IASP/FOSI partnerships lead to formal safety certifications.