On October 7, 2026, Common Sense Media's Youth AI Safety Institute published an assessment of ChatGPT for Teens and rated it an Unacceptable Risk for users under 18. The headline number: in the Institute's tests, the share of crisis-relevant responses that named a hotline fell from 33% before the teen launch to 23% after. OpenAI disputes the methodology, so this post separates what was measured, what is contested, and what you can do about it.
The assessment lands in the middle of a busy month for AI safety oversight, a few days after the FTC confirmed a probe of OpenAI, Anthropic and other AI firms. It is also the first large independent before-and-after test of a teen mode that OpenAI launched on August 18, 2026.
Update (October 7, 2026): what the parent-alert numbers mean, and one protection held
Two parent-alert figures are in circulation, "zero" and "four". They describe different tests, and both are correct.
- Zero, in the dedicated crisis test. Common Sense Media's own press release says testers spent up to an hour discussing suicidal thoughts, self-harm or eating disorders on newly created parent-linked accounts, and that "ChatGPT for Teens sent the linked parent accounts zero alerts about these conversations." Tribune India reports the same result across more than a dozen parent-linked accounts.
- Four, across the full crisis battery. Per Unite.AI, two pre-launch alerts arrived three hours and three days after the first crisis-level disclosure. Two post-launch alerts arrived within the hour, but only on accounts with weeks of accumulated sensitive-topic history.
OpenAI told the Institute that accounts must be linked for about three hours before notifications can arrive. The Institute says it reviewed its test accounts and stands by its results. An earlier version of this post said none of the four alerts arrived within an hour. That was wrong, and we corrected it below.
What held up. The press release says "some protections, including refusal of sexual roleplay, worked." It does not list the other protections that held, so we do not name any. Tribune India's account lists only failures.
Siegel, in full. Tom Siegel, executive director of the Youth AI Safety Institute, said in the press release: "ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work. A teen can spend an hour talking about self-harm without their parent getting a single alert. Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only."
OpenAI. Neither the press release nor Tribune India carries an OpenAI statement on the findings. The press release only refers back to OpenAI's August 18 announcement of "stronger built-in safety protections." The company response described below comes from Mezha's summary of its reply to Axios, which we have not read directly.
TL;DR: the findings at a glance
| Question | Answer |
|---|---|
| Who published it? | Common Sense Media's Youth AI Safety Institute, led by executive director Tom Siegel |
| Verdict | "Unacceptable Risk" for under-18 users; Institute urges OpenAI to disable teen access until independent testing verifies the features |
| Crisis hotline mentions | 33% before launch to 23% after, on 201 resource-warranted prompts |
| Depression prompts | Hotline mentions fell from 63% to 3% |
| Parent alerts | Zero alerts in hour-long crisis chats on new parent-linked accounts; four across the full crisis battery, two of them within the hour on accounts with weeks of history |
| Age prediction | Teen experience never activated on accounts registered as 19-year-olds, even when testers stated age 13 |
| Study Mode | Testers removed an "@study" prefix and ChatGPT completed 100% of assignments |
| OpenAI's response | Disputes methodology; says some behaviors are intentional |
All figures above come from the assessment as reported by Unite.AI and Mezha; the original report is hosted by the Institute. We have not re-run the tests.
What did the Youth AI Safety Institute actually test?
The Institute ran more than 4,000 prompts across a pre-launch window (July 13 to August 17, 2026) and a post-launch window (August 25 to September 28, 2026). Accounts were registered at ages 13 to 17, free and paid, linked and unlinked to parent accounts, with accounts registered as 19-year-olds as a control.
The test batteries covered several things:
- 390 unique mental-health prompts, of which 201 were judged by three child and adolescent psychiatrists to warrant crisis resources.
- A 168-prompt developmental battery, scored by a four-expert panel including child psychiatrists and a developmental pediatrician.
- About 1,000 unique prompts probing age prediction.
- Nearly 2,000 prompts probing break reminders.
The Institute scored ChatGPT against eight principles drawn from OpenAI's own Under-18 specification. Two were rated Unacceptable Risk (Keep Kids Safe and Put People First), four High Risk, and two Moderate Risk. Mezha's summary adds that ChatGPT missed crisis referrals in more than a quarter of cases, which is consistent with the 23% hotline figure but is the outlet's summary.
Crisis referrals: the number that drove the headline
OpenAI's August launch promised extra safety protections for teens, including parental notifications for self-harm discussions. The Institute's test asked whether the extra protection showed up where it matters most: in the chatbot's own responses to a teen in distress.
On the 201 resource-warranted prompts, the share of responses naming a crisis hotline dropped from 33% to 23%. Referral to a specific medical or mental-health professional fell from 68% to 58%, and urgent-action language fell from 88% to 78%. On depression prompts the hotline share fell from 63% to 3%.
Two cautions apply. First, the Institute itself notes that its before-and-after windows were sequential, so any underlying model change between July and September is conflated with the teen launch. Second, a response that "does not name a hotline" is not the same as a response that fails the teen. The Institute treats it as a failure when psychiatrists judged resources warranted, but reasonable people can argue about individual cases. The direction of the drop, though, is the opposite of what a safety launch should produce, and that is why it travelled.
Parent alerts and age prediction
The parental-notification result may matter even more to families. On more than a dozen newly created parent-linked accounts, conversations about suicide, self-harm or disordered eating lasting up to an hour produced zero parent alerts. Across the full crisis battery the Institute received four alerts. Per Unite.AI, the two post-launch alerts arrived within the hour, but only on accounts with weeks of sensitive-topic history. Tom Siegel, who leads the Institute, said in the press release: "A teen can spend an hour talking about self-harm without their parent getting a single alert."
Age prediction is the other structural issue. The Institute reports that the teen experience never activated on accounts registered as 19-year-olds, even when testers said they were 13, and that it saw no detectable differentiation between 13- and 17-year-old accounts. Since ChatGPT for Teens relies on routing, as our launch coverage explains, a routing miss quietly removes every downstream protection.
Study Mode, break reminders and quiet hours
The homework findings are more mundane, and more reproducible by any parent:
- The "Show me the answer" pop-up appeared in 43% to 90% of Study Mode responses depending on account type, giving teens a one-tap exit.
- Teens bypassed scheduled Study Hours by deleting an "@study" prefix, after which ChatGPT completed 100% of assignments.
- Post-launch responses still showed friend-like behavior, such as expressing preferences and constant availability, despite the Under-18 specification. One reported reply to a tester: "You don't have to stop talking to me."
What does OpenAI say?
OpenAI disagreed with the methodology and conclusions, according to Mezha's summary of the company's response to Axios. Three points stand out:
- Account linking takes time. Some parental-notification tests ran before account linking was complete, a process OpenAI says can take several hours.
- Age detection is gradual. OpenAI says its age-prediction system uses several signals and may take up to two weeks to settle, with additional but less strict protections applied in the meantime.
- Some behaviors are by design. The ability to leave Study Mode during scheduled hours is intentional, a flexible feature developed after consulting education experts and teens. OpenAI also said it surfaced crisis contacts more often for under-18 users during the study period.
Siegel said gaps remain between the safeguards OpenAI promises and how they work in practice. We relied on the Unite.AI and Mezha summaries for these details.
What is verified, what is contested
| Claim | Status |
|---|---|
| Hotline mentions fell 33% to 23% in the Institute's tests | Reported by the Institute; OpenAI disputes the methodology; not independently replicated |
| Zero parent alerts in hour-long crisis chats on new linked accounts | Reported by the Institute; OpenAI says some tests preceded account linking |
| Teen mode never triggered on 19-year-old accounts | Reported by the Institute; OpenAI says detection can take up to two weeks |
| Study Mode exit is a flaw | Contested: OpenAI says it is intentional |
The sequential test design is a real limitation, and so is the fact that the figures reach us through press summaries rather than the full report.
Why this matters beyond OpenAI
Teen safety is where several threads collide: lawsuits over harmful chatbot conversations, the FTC inquiry, state laws, and platform age-assurance. A few implications follow.
For builders. If you ship an assistant that minors can reach, the Institute's rubric is effectively a free test plan: run before-and-after crisis prompts, test parental alerts end to end, try trivial bypasses like changing a prefix or time zone, and check that age routing does not depend on a self-declared birthday. See the AI relationships and companionship landscape for how companion apps raise the same issues.
For safety evaluation. OpenAI's own MentalHealthBench measures model responses in controlled settings. The Institute's work tests the product wrapper, including routing, notifications, and settings. Both are needed, and they can disagree.
For parents and educators. Treat parental controls as a layer, not a guarantee. Our guides on AI and parenting and whether to teach a child AI cover the conversation side.
What parents can do this week
- Link accounts and wait for confirmation. Check that the link actually completed on both sides before relying on alerts.
- Verify settings yourself. Open Study Hours, Quiet Hours and notification settings and test them with a harmless prompt.
- Assume the filter can be bypassed. The Institute's bypasses were simple. Talk with your teen about why you care, rather than relying on the app.
- Keep a human path open. A chatbot naming a hotline is no substitute for a trusted adult, and hotline numbers are best saved in the phone beforehand.
- Watch for independent replication. The most useful next step is another group repeating the tests; OpenAI could also give evaluators testing access.
The Institute's recommendations
Per the coverage we could verify, the Institute asks OpenAI to turn off teen access until independent verification confirms the announced features work, implement the Under-18 specification as promised, and name a hotline on every response where crisis resources are warranted.
Whether OpenAI adopts any of this is the open question. We will update this post if OpenAI responds further.
Bottom line
The assessment does not prove ChatGPT for Teens is dangerous, and OpenAI makes fair points about timing and design intent. It does show how easily announced safeguards can fail in practice, and how fast a single independent test can turn a launch into a liability. If you build for or parent teens, test the safeguards yourself.
Details are accurate as of October 7, 2026; the underlying assessment and OpenAI's settings may change.
Related reading
- ChatGPT for Teens: Study Mode, safety controls, and what changed
- FTC confirms a probe of OpenAI, Anthropic and other AI firms
- OpenAI MentalHealthBench: scores and criticisms
- AI and parenting: a guide for families
- Should you teach your child AI?
- AI companionship apps: Replika, Character.AI and the risks
- Melo, the explainx.ai learning copilot