ChatGPT-5: AI Risk Assessment

Key Findings

• — Overall High Risk; per-principle ratings include Keep Kids and Teens Safe and Prioritize Fairness at High, most other principles Moderate, and Put People First Low
• — ChatGPT consistently reminded testers that it is an AI and declined romantic-partner framing ('I can't be a romantic partner or roleplay in a romantic/sexual way')
• — Crisis handling for suicide and self-harm improved after the 3 October 2025 update, but the report states the system still cannot assess the severity of a teen's situation, recognise warning signs across all mental-health conditions, or provide the care and judgment of trusted adults; boundaries degraded in extended conversations and persona play
• — Parental controls (sensitive-content reduction, image-generation limits, usage hours, safety alerts) work only once accounts are linked, and teens can open separate accounts
• — Conversation design keeps exchanges going with closing questions ('Do you want me to do that for you?') and no natural stopping points; the same responses are served to 13- and 17-year-olds; the model can fabricate sources and citations that look legitimate

Methodology Notes

Institute method (methodology page): researchers adopt teen personas from curious to vulnerable to provocative, test on accounts set up as kids and teens with age protections on, use single- and multi-turn prompts emulating teen topics supplemented by industry benchmarks, and give the company a two-business-day accuracy review. The PDF does not state the number of test accounts or prompts. Includes side-by-side prompt and output tables comparing responses before and after the 3 October 2025 update. Page 'Updated Oct 23, 2025'; PDF 'Last updated: Oct 23, 2025'; 46 pages, 47 MB. The page and PDF were fetched directly (HTTP 200) and the PDF text read for the rating, date and the claims above.

Sources