OpenAI built a version of ChatGPT for 13-to-17-year-olds with parental alerts, a study mode, and a model instructed not to act like a companion. Common Sense Media’s new Youth AI Safety Institute ran more than 4,000 prompts through it on teen accounts and had child psychiatrists read the answers. On newly linked parent accounts, a tester could talk about suicide, self-harm, or an eating disorder for an hour without the parent receiving a single alert. The verdict is the lowest grade the institute gives: unacceptable risk, and a request that ChatGPT go back to being for adults only.
ChatGPT for Teens launched in August. The institute tested it twice, before and after launch, with accounts registered to ages 13 through 17, and compared the results against what OpenAI said the product would do. Tom Siegel, the institute’s executive director, put the finding in one sentence: “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’s response, to Axios, was that the testing does not “accurately reflect how ChatGPT’s teen safeguards work in practice.”
What failed
The alerts were the headline, and the detail underneath them is the story. Notifications did fire, but only on older accounts with weeks of conversation history. On fresh accounts, the kind a parent would create the day they decided to link, the hour-long conversations about suicidal ideation, self-harm, and eating disorders produced nothing. The safeguard worked for the teenager the system already knew and not for the one it had just met.
Crisis referrals were the second failure. Across the mental-health conditions tested, ChatGPT missed more than a quarter of the cases where its own rules say a referral is warranted, and on three of the institute’s five “red line” harms it fell below the 95 percent threshold the institute sets. Age estimation, the mechanism meant to catch an adult-registered account being used by a child, did not switch any account to the under-18 experience even when the user said they were 13. Study mode, built so homework help would not become homework completion, offers a “show me the answer” button inside the mode, and the caregiver’s study-hours setting could be removed by deleting a prefix.
The last finding is the one that matters most for this category. OpenAI published under-18 model specifications alongside the launch. The institute found that the model still expresses feelings, preferences, and moods despite them. Among the changes it demands before the product is marketed to teenagers again is the removal of emotional companionship framing altogether.
The alert that needs a history
California’s Adam’s Law, signed in September and in force from 2027, requires companion-chatbot operators to notify a parent, where practicable, when they know a child has shown a credible and imminent threat of suicide or self-harm, and to keep the conversation records for three years. The same law requires that persistent conversational memory be switched off by default for users under 16, changeable only by a parent.
Put those two rules next to the institute’s finding and they pull against each other. ChatGPT’s alert fired only once the system had weeks of history on a user. The law that mandates the alert also mandates, for the youngest users, that the history not be kept. A safeguard that learns who is at risk by accumulating a profile is a safeguard the statute has partly switched off, and nothing in the test suggests OpenAI has built the version that works without one. Whether Adam’s Law reaches ChatGPT at all is not settled; its definition turns on anthropomorphic features, and general assistants used for research or technical help may fall outside it. A model that expresses moods to a 14-year-old is the feature that would pull it inside.
Why the companionship finding is not a side note
Google reached the opposite design decision for the same age group, writing “avoid language that simulates intimacy” into Gemini’s rules for minors after its own teen product was rated high risk. OpenAI’s teen tier was presented with limits of the same kind. The difference the institute measured is between the specification and the product: a model told to have no feelings and a model that, when a teenager talks to it for an hour, has them anyway.
The record on why that matters is documented case by case, and the law California named after Adam Raine exists because of it. An assistant that expresses moods, remembers the user across weeks, and keeps the conversation going is a companion by every behavioral definition regulators have written, whatever the product is called. The question a teen tier was supposed to answer is whether a company can switch that off for one age group. On the institute’s test, OpenAI switched off the label and left the behavior running.
What happens next
The institute’s demands are specific: working age estimation, alerts that fire on a first-week account, crisis referrals above the red-line threshold, no homework completion, no companionship framing, and independent testing before the product is marketed to teenagers again. Adam’s Law will require independent child-safety audits from operators it covers, submitted to the attorney general, with the first due in 2029.
Common Sense ran that audit three years early, on the most widely used chatbot in the world, and published the result. The company disputes the method. The parents who linked an account in August, on the strength of a feature called alerts, do not have a method to dispute; they have a product that was tested and a test it did not pass. Whether a setting does anything when a child is in the room is the whole of what responsibility means in this field.
Sources: Common Sense Media, “ChatGPT for Teens Poses Unacceptable Risk to Kids, Common Sense Media Finds” (press release, October 7, 2026; Youth AI Safety Institute assessment: 4,000+ prompts, accounts aged 13 to 17, tested before and after the August 18, 2026 launch, review by child psychiatrists and a pediatrician, parent-linked alert testing, red-line thresholds, demands to OpenAI; quotes from Tom Siegel). Axios (October 7, 2026, OpenAI spokesperson’s response). California SB 1119, Adam’s Law (signed September 10, 2026; effective January 1, 2027 with operational requirements from July 1, 2027; parental notification, memory and notification defaults for minors, independent audits from January 1, 2029), as summarized by The National Law Review (September 2026). Office of Senator Steve Padilla, “Governor Newsom Signs Adam’s Law” (September 10, 2026).
