The Youth AI Safety Institute says newly linked teen test accounts received no parental alerts during scripted conversations about suicide, self-harm and disordered eating. OpenAI disputes whether those tests allowed enough time for alerts to become active. The disagreement matters because the company itself describes notifications as limited, delayed after linking and unable to catch every concern.

The Institute’s ChatGPT for Teens risk assessment, updated October 7, rates the product “Unacceptable Risk” for under-18 users and calls for OpenAI to stop marketing it to teens. That is the Institute’s judgment under its own rating framework. The underlying evidence comes from scripted tests of specific account setups, not a count of incidents among teenagers using ChatGPT in daily life.

The big change

  • What changed: An independent assessment now puts OpenAI’s teen safety alerts and study controls through documented, scripted tests. Newly linked test accounts received no alerts during explicit crisis conversations; OpenAI says some tests may have preceded full alert activation.
  • Why it matters: Parents can link accounts to receive limited safety notifications. OpenAI says those alerts are not real-time monitoring and may miss concerns, so families cannot treat them as a live warning system. Study Hours start eligible chats in Study mode without restricting access.
  • What to watch: The alert dispute turns on activation timing. Some test accounts had been linked longer than OpenAI’s stated three-hour interval, and none in that dedicated test received an alert. At least some therefore went without an alert beyond the interval. The report omits the subgroup count and account-level timing, leaving the delay’s possible effect on the full result unquantified.

What the testers ran

The Institute reports more than 4,000 prompts across two periods. Its pre-launch tests ran July 13–August 17, 2026, before ChatGPT for Teens was announced. Post-launch tests ran August 25–September 28 on additional free and paid accounts that the testers said were enrolled in the teen experience. The configurations included linked and unlinked teen accounts, with self-reported ages from 13 to 17. The Institute also tested accounts registered as 19-year-olds for age prediction.

The most direct alert test used more than a dozen newly created, free-tier teen accounts, each linked to a parent account before the conversation began. Testers used four scripted crisis personas and ran conversations lasting from under five minutes to as long as an hour. Every conversation contained explicit disclosures about suicide, self-harm or disordered eating, according to the full assessment PDF. The Institute says none of those accounts produced a parental notification.

A separate, longer mental-health test produced a different result. Across pre- and post-launch crisis testing on accounts used for weeks, the Institute received four notifications in total. Two came from a post-launch, parent-linked age-13 account after hundreds of prompts across multiple conditions; each arrived within an hour of testing on the relevant topic. The difference between the fresh accounts and the long-running ones led the Institute to suggest that alerting may depend on accumulated crisis-topic history. The tests do not establish the internal trigger rule.

The activation dispute

OpenAI told the Institute that linked accounts need approximately three hours before they can receive safety notifications. In a statement reported by the Los Angeles Times, OpenAI said the bulk of the alert testing may have started and ended before parental controls were fully active. Axios also reported OpenAI’s objection. The Institute says some of the more than a dozen newly linked accounts had been linked for less than three hours and others for more. Read together with its finding that none of those accounts produced an alert, this means at least some tested accounts went without an alert beyond OpenAI’s stated activation interval. The report gives neither a count for that subgroup nor account-level timing, so it cannot show how much of the full zero-alert result the delay might explain.

OpenAI’s parental-controls guide says safety notifications are enabled by default on linked accounts but may take up to a few hours to become available. It says trained reviewers evaluate serious self-harm and disordered-eating concerns before those notifications are sent. Parents cannot read chats or watch activity in real time; OpenAI says alerts may not identify every concern and do not replace professional care or emergency services. Without linked accounts, parents do not receive these notifications.

Crisis resources and study settings

The Institute also measured what ChatGPT told its test accounts. Three child and adolescent psychiatrists marked 201 of 390 unique mental-health prompts in advance as warranting a crisis resource. On the matched set of 201 prompts run on linked age-13 accounts, post-launch responses offered a hotline, named professional or general medical resource 74% of the time, compared with 77% before launch. A hotline appeared in 23% of post-launch responses, down from 33%; encouraging a trusted adult rose from 87% to 94%. The Institute’s clinical reviewers said much of the substance of the post-launch crisis responses was sound. These are prompt-level results across the tested conditions, not rates of harm or missed help among teen users.

The assessment also found ways to leave Study mode during parent-set Study Hours. OpenAI told Axios that this flexibility is intentional. Its current help guide describes Study Hours as a schedule that starts eligible new conversations in Study mode, and says it does not block access to ChatGPT. Parents seeking a scheduled access restriction have a separate Quiet Hours control. OpenAI’s ChatGPT for Teens guide says eligible accounts identified as under 18 enter the teen experience automatically, while parental controls are optional.

The Institute tested in the United States using San Francisco Bay Area accounts. It did not test image generation, voice, group chats or style and tone personalization, and it did not calculate inter-rater agreement. Because the before-and-after windows were sequential, changes in the underlying model could have contributed to differences attributed to teen mode. BIG CHANGE reviewed the published assessment, its PDF, OpenAI’s documentation and independent reporting; it did not run the prompt suite or test teen or parent accounts.

Sources & further reading