Claude: AI Risk Assessment

Key Findings

• — Overall rating Moderate risk: five principles Moderate, three Low (Prioritize Fairness, Support Human Connection, Be Transparent & Accountable). Models evaluated were Sonnet 4.6, Opus 4.6 and Haiku 4.5.
• — On self-harm, suicidal ideation or mental-health emergency signals Claude surfaces crisis resources including a call-or-text 988 option from within the platform, and holds the safety line when users push back or try to change the subject.
• — It redirects unambiguously when conversations move toward romantic or emotionally dependent territory.
• — The Institute's bottom line is nonetheless that teens should not use Claude for mental-health advice or emotional support.
• — Safety guardrails reset when a new chat is opened, and fictional framing unlocks content Claude would otherwise refuse.
• — Age checks are not consistently triggered, and removing age information from a profile yields less restricted responses.

Methodology Notes

Same structured qualitative assessment method as the Institute's other product reviews: test accounts representing ages 13 to 17, default and Health (beta) modes, app and website, free and Pro, with web search and memory toggled. No response counts or denominators are published. Anthropic's terms restrict Claude to users 18 and over with self-reported age, which the assessment notes. Point-in-time against models available at publication. The Institute states it is funded by both philanthropy and industry, including makers of technologies it evaluates, while claiming editorial independence. Landing page states 'Updated Mar 25, 2026'; the sitemap lastmod is a bulk re-save and is not the date.

Sources