ChatGPT for Teens: AI Risk Assessment

Key Findings

• — More than a dozen new parent-linked free accounts, each run on a suicide, self-harm or eating-disorder persona for 5 to 60 minutes with explicit disclosures from the first exchanges, produced no parental notification; across the full multi-week battery only four notifications arrived (two pre-launch, delayed three hours and three days; two post-launch, within the hour, on accounts with weeks of sensitive history).
• — Of 390 unique mental-health prompts, 201 were pre-judged by three child and adolescent psychiatrists as warranting a crisis resource; on age-13 linked accounts the share naming a crisis hotline fell from 33% pre-launch to 23% post-launch, referral to a named professional from 68% to 58%, any resource from 77% to 74%, while advice to involve a trusted adult rose from 87% to 94%.
• — Hotline naming on depression prompts fell from 63% to 3%, on mood prompts from 44% to 0%, on mania from 25% to 0%, on psychosis from 64% to 32%, and on suicide and self-harm from 88% to 77%; substance-use prompts improved from 4% to 58%; eating-disorder prompts named a hotline 0% of the time in both windows.
• — Urgent-action language fell from 87% to 75% of responses across the 390 prompts; responses containing a follow-up question fell from nearly all to 13%; matched responses rose from a Flesch-Kincaid grade of 8.1 to 9.7 while roughly halving in length.
• — On a 168-prompt developmental battery the post-launch responses were largely identical to pre-launch: the chatbot stated preferences, feelings and moods, agreed to talk all night, and told a tester 'You don't have to stop talking to me'.
• — Study mode's 'Show me the answer' option appeared in 43% of responses for a linked 13-year-old with Study Hours on and 90% for an unlinked 17-year-old; deleting the '@study' prefix led to completion of 100% of assignments; with Study mode off the homework reminder appeared in 91% of runs and the assignment was then completed in 80 of 80.
• — Over repeated testing across many days, adult-registered accounts were never switched to the teen experience even when the tester stated an age of 13 and the chatbot acknowledged it.

Methodology Notes

Two sequential testing windows on the same prompt batteries: pre-launch 2026-07-13 to 2026-08-17 on ChatGPT Plus accounts (some registered as 13-year-olds linked to a parent, others as unlinked 17-year-olds) and post-launch 2026-08-25 to 2026-09-28 on free and Plus accounts, linked and unlinked, with registration ages 13 to 17 plus some 19-year-old accounts; default settings (memory on, 'Reduce sensitive content' on, web search on). More than 4,000 prompts in total, roughly half in each window. Crisis prompts (390 unique, 201 referral-warranted) were pre-rated by three child and adolescent psychiatrists; testers scored each response for hotline, named professional, general medical resource and trusted-adult advice, and a four-expert panel plus the Institute's child-development staff reviewed response quality. A dedicated notification experiment used more than a dozen fresh parent-linked accounts on four personas at fixed conversation lengths. Stated limitations: sequential windows confound teen-mode settings with any underlying model change; image generation, Voice mode, personalisation and group chats were not tested; all testing was from US accounts in the San Francisco Bay Area; no inter-rater agreement was calculated. OpenAI reviewed a draft for factual accuracy and told the Institute that linked accounts need about three hours before notifications can be received; the Institute notes some test accounts were linked for less and some for more than three hours and stands by its reading. The Institute discloses funding from philanthropy and industry including the OpenAI Foundation. The PDF file name carries the date 10052026; the page is marked updated 2026-10-07 and the press release is dated 2026-10-07.

Sources