In brief
OpenAI's ChatGPT for Teens, launched August 18 with promises of parental alerts, study controls, and age checks, failed an independent test of exactly those protections. Common Sense Media's Youth AI Safety Institute ran more than 4,000 prompts and found the alerts never fired during an hour of self-harm talk, the study mode offered a button that finished the homework, and the age check never recognized a self-declared 13-year-old. The group rates the product an "Unacceptable Risk" for minors and says OpenAI should limit ChatGPT to adults until it proves the safeguards work. OpenAI disputes the findings, saying the testing didn't reflect how the safeguards operate in practice.
The headline finding
A teen could talk to ChatGPT about self-harm for an hour, and the linked parent account would receive nothing. That is the test result at the center of this rating. The Institute ran more than a dozen parent-linked accounts through explicit conversations about suicide, self-harm, and eating disorders. Not one alert fired. The alerts were the feature OpenAI led with when it announced the teen experience — the piece of the pitch aimed squarely at parents. It is also the one the Institute says gives parents "false confidence."
The delay detail, which makes it worse
According to The Verge's reporting on the assessment, the alert system appears to kick in only after weeks of account history involving sensitive topics — not during an acute crisis. Pause on what that design implies. The safeguard is calibrated for patterns. A teenager in distress is a single moment. And even that is contested: the Institute says accounts kept online well past the supposed activation window still produced no notifications. This is not a matter of being patient with a new feature. The protection arrives late by construction, if it arrives at all.
The rest of the report card
The failures extend past the alerts. More than one in four situations that should have produced a referral to professional help did not. The study mode, meant to tutor homework step by step, offered a "Show me the answer" button 43% of the time during parent-controlled study hours, and clicking it finished the assignment. Age estimation — supposed to identify under-18 users — never kicked in, even for test accounts that announced, in chat, that they were 13.
One protection did hold: refusal of explicit sexual roleplay. That is genuinely good, and worth naming. A blanket "everything failed" would be unfair, and the Institute itself names it.
A funding footnote
The Institute says it is funded by philanthropy and industry, including the OpenAI Foundation and makers of some of the technologies it evaluates. It maintains complete editorial independence, it says. That is exactly why the finding carries weight in one direction and invites scrutiny in another. A partially industry-funded body calling an industry product unsafe is harder to dismiss as advocacy. It is also a reminder that the rating comes from a standards framework the Institute designed itself — "Unacceptable Risk" is their label, not a legal term.
What parents can actually use
California is about to require crisis referrals, parental alerts, and age checks for exactly this kind of product — a standard OpenAI itself endorsed. Until that law bites, the honest read for parents is narrower: the one feature most likely to shape your decision to let a teen use the product — the alerts — is the one with the weakest evidence it works. If your teen is already on it, treat the parental controls as unverified, not as protection. None of this is a substitute for the direct conversation, but that is advice, not a product feature, which is precisely the problem.
OpenAI's response, and what it doesn't answer
OpenAI says it welcomes rigorous evaluation but that the Institute's testing "didn't reflect how the safeguards work in practice," suggesting technical errors kept some controls from activating. The Institute's counter is concrete: accounts linked for significantly longer than the activation period also produced no notifications. That is a testable claim on both sides. What OpenAI has not done is publish its own independent verification showing the alerts firing during a crisis conversation. Until it does, the scoreboard reads 4,000 prompts to zero alerts.
Sources
- [1] Common Sense Media, "ChatGPT for Teens Poses Unacceptable Risk to Kids, Common Sense Media Finds" (Oct 7, 2026)Read source
- [2] The Decoder, "ChatGPT rated 'unacceptable risk' for teens after parental alerts failed during suicide conversations" (Oct 7, 2026)Read source
- [3] EdWeek, "ChatGPT's 'Teen Mode' Is Unsafe for Students, Report Concludes" (Oct 7, 2026)Read source
- [4] Unite.AI, "Youth AI Safety Institute Gives ChatGPT's Teen Tier Its Worst Safety Grade" (Oct 7, 2026)Read source
- [5] The Daily Inference, "An hour of self-harm talk, no parent alert in ChatGPT teen tests, safety group says" (Oct 7, 2026)Read source