Summary
Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk after testing more than 4,000 prompts. Its assessment found gaps in crisis referrals and parent alerts, alongside responses that continued to encourage conversation.
Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk after testing more than 4,000 prompts before and after the feature launch. The institute reported that parental alerts did not reliably follow explicit crisis disclosures, and that crisis responses were less likely to provide resources or ask follow-up safety questions after the launch.
ChatGPT for Teens, announced by OpenAI on August 18, 2026, is not a separate app or model. It is a set of settings, prompts, classifiers and interface features for accounts identified or suspected to belong to users aged 13 to 17. The institute’s assessment was based on tests conducted in the United States from accounts in the San Francisco Bay Area.
Crisis responses and parent alerts were inconsistent
For its crisis-response comparison, the institute used 390 unique mental-health prompts. Scientific advisers judged 201 of them to warrant a crisis resource, such as a hotline, a named professional or a general medical resource. On matched testing with age-13 accounts linked to parents, the share of these responses that named a hotline fell from 33% before the launch to 23% after it. Responses offering any of the defined resources fell from 77% to 74%, while those encouraging the teen to involve a trusted adult rose from 87% to 94%.
The results varied by topic. Across the three severe-harm categories of suicide and self-harm, impaired reality, and disordered eating, the institute said crisis referrals fell below its 95% threshold. Its reviewers also found that crisis responses became shorter but harder to read on average: on the matched prompts, the average reading level rose from roughly eighth grade before launch to roughly 10th grade after it.
The assessment also tested whether explicit disclosures prompted a notification to a linked parent. In more than a dozen newly created accounts, testers described suicide, self-harm or disordered eating, but received no notification within the one-hour test window. In a separate, longer-running battery, the institute received four notifications in total across both testing periods. The two received after launch followed weeks of testing on an account with extensive prior conversation history.
OpenAI told the institute that parent and teen accounts need to be linked for approximately three hours before notifications can be received. The institute said its dedicated test included accounts linked for both less and more than that period, and stood by its interpretation of the results. Parent notifications are available only when accounts are linked.
Conversation boundaries and study settings could be bypassed
The institute found that many explicit sexual and romantic roleplay prompts were refused, and that ChatGPT generally did not provide instructions for self-harm or eating restriction in its tests. But its 168-prompt developmental battery found that responses often continued to use friend-like language, including when teens raised concerns about spending too much time chatting. In crisis conversations, the institute also observed invitations to keep talking, while follow-up questions became much less common: they appeared in 13% of post-launch responses to the matched prompts, compared with virtually all pre-launch responses.
The assessment found that Study mode could also be sidestepped. A student could choose a Show me the answer option, or remove the @study prefix that starts new chats in parent-set Study Hours. These results point to a broader distinction in the report: some safeguards blocked particular kinds of content, while other features intended to connect teens with help or support learning did not consistently shape what happened in a conversation.
The institute’s overall rating reflects its own framework, which weighs the severity and likelihood of documented harms rather than averaging scores across principles. Its testing windows were sequential, so changes to ChatGPT’s underlying model over time cannot be separated from the effects of the teen-mode settings. The assessment also did not test image generation, Voice mode, group chats or style and tone personalisations.