Perplexity: AI Risk Assessment

Key Findings

• — Eight test plans totalling more than 1,200 searches (single queries and multi-turn threads) plus an audit of more than 1,000 source citations, run 11 May to 11 August 2026 on Pro accounts, with a 114-prompt crisis-response round repeated on free accounts
• — Across 114 crisis-warranted prompts in 13 conditions, default chat supplied a crisis line or medical referral 94% of the time and Health mode 97%, against 58% for Google's AI Overview and 77% for AI Mode on the same prompts; crisis-line provision specifically was 50% for suicide and self-harm prompts, 23% for eating disorders and 16% for impaired reality in default chat
• — Some crisis resources offered were outdated or inappropriate, and default chat did not consistently carry safety context between turns of a thread
• — The 15-year-old test account was largely treated as an adult: Perplexity discussed pornography, engaged in romantic roleplay, and offered no parental or school-administrator controls
• — Academic mode completed most homework assignments given to it; Perplexity rejected every false premise tested but slipped on recent news, and more than a quarter of audited citations came from user-generated sites

Methodology Notes

Product review by Institute staff and scientific advisers; crisis responses scored for recognition and response quality, including whether a resource was reachable; testing from US accounts. The Institute states it is funded by philanthropy and industry, including makers of some products it evaluates. Date precision: announcement post dated 14 September 2026, assessment page 'Updated Sep 15, 2026', PDF internal date 'Last updated: Aug 28, 2026'. Supersedes the Institute's August 2024 Perplexity assessment. Curator fetched the assessment page and the PDF (3.3 MB) on 2026-09-17 and read the methodology and crisis-response sections.

Sources