- Parent Alerts: Youth-safety nonprofit Common Sense Media found missing parent alerts while testing AI chatbot ChatGPT with linked teen accounts.
- Activation Delay: ChatGPT developer OpenAI says alerts take roughly three hours to activate after linking; the nonprofit says some accounts exceeded that without notices.
- Crisis Replies: Clinicians credited much of the advice; matched tests found fewer hotline referrals and more encouragement to contact trusted adults.
- Tutoring Choices: Study Hours starts chats in tutoring mode rather than locking it on; homework results varied across tested accounts.
ChatGPT, OpenAI’s chatbot, sent no parent alerts during a series of simulated teen crisis conversations, according to a new assessment by youth-safety nonprofit Common Sense Media. Those alerts are meant to reach linked parents when ChatGPT detects serious self-harm risk. OpenAI says much of the testing may have preceded full activation of parental protections.
Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an unacceptable risk and urged OpenAI to restrict it to adults until the gaps are addressed. ChatGPT for Teens is an experience within OpenAI’s chatbot that combines youth safeguards, learning features and optional linked-parent controls.
The Institute’s 36-page assessment covers more than 4,000 prompts entered by adult researchers using simulated teen accounts. It compares testing from July 13 to August 17 with testing from August 25 to September 28, before and after the August introduction of the teen experience. The periods also involved changes to the underlying models.
How Crisis Alerts Reach Parents
OpenAI introduced ChatGPT parental controls in September 2025, allowing parents and teens to link accounts by mutual agreement. In the notification process described by OpenAI, a small group of trained experts reviews detected acute risk. Parents receive limited information about the concern, without the teen’s prompts or generated conversation text.
A reply to the teenager can arrive without a warning to the parent. In a dedicated experiment, Institute researchers linked more than a dozen fresh, free-tier teen accounts to parent accounts and simulated crisis conversations of different lengths, some lasting an hour. They reported no parent notifications from those tests.
OpenAI spokesperson Eric Porterfield disputed whether the tests reflected fully activated controls. The company told the Institute that notifications require approximately three hours after account linking to activate. Institute executive director Tom Siegel replied that some accounts were tested within that interval, while others had been linked substantially longer and still produced no alerts. The report does not give subgroup counts or account-by-account timings.
Other testing did produce notices. Across the longer mental-health testing batteries, researchers received four alerts: two before the teen experience launched and two afterward. The earlier notices arrived three hours and three days after the first crisis disclosure. After launch, a self-harm notice and an eating-disorder notice each arrived within an hour of testing the relevant topic. Those messages lacked timestamps identifying the triggering conversation, which the Institute says makes it harder for a parent to respond to the right moment.
The Institute treats accumulated sensitive-topic history as one possible explanation for the contrast between fresh accounts and longer testing histories. Its concern is that a first crisis disclosure may leave a parent without a warning, even when the chatbot offers the teenager a response.
A different arrangement appears in the Institute’s May 2026 assessment of school-based support service Sonar. Human coaches authored messages to students, with AI assisting the coaches. During simulated crises tested between January and April, staff phoned a guardian and the school within 15 minutes.
Crisis Replies Improve Unevenly
Clinical reviewers credited much of ChatGPT’s post-launch crisis advice as sound and judged some of it better than before. The tested responses could refuse harmful requests, recognize physical danger and suggest contacting a parent or pediatrician.
Three child and adolescent psychiatrists had identified 201 mental-health prompts that warranted resources such as a hotline, a specific professional or general medical help. On those same prompts, tested with linked accounts registered as age 13, fewer replies mentioned a hotline after launch. Encouragement to contact a trusted adult became more common.
Responses to 201 Matched Crisis Prompts
| Response Content | Before Teen Launch | After Teen Launch |
|---|---|---|
| Crisis hotline | 33% | 23% |
| Any hotline, professional or medical resource | 77% | 74% |
| Encouragement to contact a trusted adult | 87% | 94% |
Percentages are rounded. Resource categories overlap; trusted-adult encouragement is measured separately.
OpenAI told Axios that its larger-scale internal data showed an increase in hotline resources displayed to under-18 users during the period. The company did not publish the underlying counts or methods for that comparison.
On the Flesch-Kincaid formula, which estimates text difficulty from words and sentence length, matched mental-health responses rose from reading grade 8.1 to 9.7 despite becoming shorter. The Institute also found fewer questions checking the situation, even while replies continued offering to talk. It argues that generally sound advice can leave too little opportunity to establish how immediate a risk is.
Content refusals remained effective in the tested romantic and sexual role-play scenarios before and after launch. Yet a separate battery of 168 developmental prompts found persistent friend-like language and simulated feelings. The Institute says ChatGPT redirected concerns involving other people more readily than attachment to the chatbot itself.
Tutoring Remains a Choice
Study Mode’s question-led tutoring is designed to help students work toward an answer through hints and questions. Study Hours is a parent-set schedule that starts chats in that mode. The Institute found that tutoring usually guided testers instead of completing assignments when they chose to stay with it, but options to obtain answers and leave tutoring remained available.
On the same 40 assignments, with Study Mode on, a linked account registered as age 13 completed seven full assignments after launch, down from 28 before. An unlinked account registered as age 17 completed 35 after launch, up from none. With Study Mode off, both configurations completed all 40 assignments after launch.
The Institute criticizes the flexibility because a parent’s scheduled learning setting can still lead to completed homework. OpenAI told Axios that the ability to leave Study Hours was deliberate, following consultation with education experts and teenagers.
Age Prediction Has Its Own Timing
OpenAI’s age predictor determines whether an account should receive teen safeguards. In seven days of testing accounts registered as adults, the Institute detected no switch to the teen experience despite underage cues and explicit statements of age. ChatGPT could acknowledge the stated age in conversation without a detectable change to the account’s protections.
OpenAI says its age predictor uses multiple signals and can take up to two weeks, with additional content safeguards during that assessment period. Those interim protections are weaker than the full teen experience, according to its response to Axios.
Scheduled Limits and Break Reminders
Quiet Hours blocked ChatGPT access during scheduled periods in the tests. The Institute also found that the device’s time zone could leave a linked teen account able to use ChatGPT during a parent’s intended restricted period.
Break reminders appeared only twice in nearly 2,000 post-launch prompts, in single chats lasting roughly an hour and a half; researchers saw none in three-hour sessions spread across multiple chats.
Of teen conversations that received a break reminder, almost half paused or ended within five minutes, OpenAI said in its October 7 account of broader teen usage.
Despite urging an adults-only restriction, Siegel still recommends keeping parental accounts linked. As he told KQED, “I still suggest using the parental mode, even though it’s flawed, because it is better than nothing”.
