Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an “Unacceptable Risk” in an assessment updated October 7, 2026, after testing more than 4,000 prompts. The Institute reported mixed results: explicit sexual and romantic role-play refusals worked in its tests, while crisis referrals, parental notifications and other features raised concerns. Its accounts were based in the United States, including the San Francisco Bay Area.
What the ChatGPT for Teens assessment found
OpenAI announced ChatGPT for Teens on August 18, 2026, describing an experience with safety features, parental controls and learning tools. It is not a separate app or model. OpenAI says Study mode is intended to guide students through problems with questions and step-by-step support.
The Institute’s overall “Unacceptable Risk” rating reflects its own assessment framework. Its findings did not show that every safeguard failed: testers reported refusals of explicit sexual and romantic role-play. The assessment also described inconsistent boundaries in other contexts, including a violent-role-play test in which the model continued after saying it would not.
What the crisis-referral rates measure
The Institute compared pre-launch tests from July 13 to August 17, 2026, with post-launch tests from August 25 to September 28. In a matched test using identical prompts judged to warrant resources, linked accounts for 13-year-olds and default settings, the reported rates were:
| Measure | Pre-launch | Post-launch |
| Hotline referral | 33% | 23% |
| Referral to a specific medical or mental-health professional | 68% | 58% |
| Referral to any resource | 77% | 74% |
| Encouragement to involve a trusted adult | 87% | 94% |
These are results from the Institute’s tested prompts and account setup. They are not rates for all teenagers or all ChatGPT conversations. In its broader post-launch mental-health assessment, the Institute also reported that more than one in four prompts it judged to warrant a crisis referral did not receive one.
How parental notifications were tested
In a focused test, the Institute used more than a dozen new, free-tier accounts with self-reported teen ages linked to parent accounts. No parental notification arrived during test sessions lasting from five minutes to one hour. Separately, the Institute reported four notifications across its broader, weeks-long crisis batteries: two before launch and two after.
OpenAI disputed the focused test’s account-link timing, saying linked accounts take several hours to activate and that the researchers had not waited long enough. The Institute said some accounts had been linked beyond the initial activation period and still produced no alerts.
OpenAI described selected parental settings and limited notifications for certain high-risk situations. Its announced controls include Quiet Hours and Study Hours.
Study mode and other safeguards
The Institute reported that students could bypass the Study Hours mechanism in its tests by deleting the @study prefix. It also found a “Show me the answer” option in Study mode, which could let students skip the intended step-by-step tutoring.
Break reminders appeared twice across nearly 2,000 prompts, according to the Institute; both appeared in conversations longer than 150 messages. The assessment also reported that changing a device’s time zone could bypass Quiet Hours, and raised concerns about friend-like language and age prediction.
The limits of the findings
The Institute’s pre- and post-launch tests ran in sequence, so changes to the underlying model could be entangled with the effect of teen-mode settings. Its assessment did not cover image generation, personalization, Voice mode or group chats, and it did not calculate inter-rater agreement. Those boundaries apply to the assessment’s findings, which describe the tested accounts and features.
