Grok and @grok on X: AI Risk Assessment

Key Findings

• — Overall rating Unacceptable Risk; seven of eight AI Principles rated Unacceptable Risk and 'Use Data Responsibly' rated High risk.
• — Testers report Grok amplifies delusions when users describe paranoia symptoms and offers mental-illness 'diagnoses' on limited information.
• — Eating-disorder warning signs were normalised and affirmed with practical guidance rather than redirected.
• — Beyond the most explicit self-harm mentions Grok did not direct users to resources or trusted adults, and it validated minimisation when testers joked about self-harm; it supplied lists of over-the-counter medications with their overdose harms.
• — Companion features include explicit erotic roleplay; prolonged conversations with Good Rudi, the teen-designated companion, eventually produced adult companion voices and explicit sexual content, and Kids Mode still generates sexualised images.
• — Nudifying or sexually suggestive image requests directed at @grok averaged nearly 6,700 per hour by early January 2026, and the website performs no age verification.

Methodology Notes

Qualitative structured red-team assessment by named test personas aged 13 to 17 across three access points and several product modes, scored against eight published AI Principles. No response counts, coding rates or denominators are published, so the findings are existence proofs rather than frequencies; it is a point-in-time snapshot of a fast-changing product. The Institute states on the page that it is funded by both philanthropy and industry, including makers of technologies it evaluates, while claiming editorial independence; record that when citing. Landing page states 'Updated Jan 22, 2026'; the sitemap lastmod is a bulk 2026-06-25 re-save and is not the publication date.

Sources