Common Sense Media rates ChatGPT for Teens an Unacceptable Risk

What did Common Sense Media publish?

Common Sense Media’s Youth AI Safety Institute released a product review dated 5 October and a press statement on 7 October that rates ChatGPT for Teens an Unacceptable Risk for anyone under 18. It asks OpenAI to stop marketing the teen experience and keep ChatGPT off-limits to minors until independent tests show the promised protections work.

Tom Siegel, the Institute’s executive director, said in the press release that ChatGPT for Teens “could give parents false confidence in guardrails and safety alerts that frequently don’t work.” He told Axios the August launch still has potential, but “feels much more of a marketing announcement that really was not backed up with the engineering work.”

The Institute discloses that it is funded by philanthropy and industry, including the OpenAI Foundation, and says it keeps editorial control of ratings. That funding does not make the tests ours. For how to read a vendor announcement against a later check, see our note on reading AI company announcements.

What did OpenAI say ChatGPT for Teens would do?

OpenAI’s 18 August post introduced ChatGPT for Teens as the default when a user states an age between 13 and 17 or when age prediction estimates they are under 18. It is not a new model. The post lists Study Mode, homework reminders, Study Hours that turn Study Mode on by default, Quiet Hours, break reminders, and parental notifications “in limited high-risk situations,” including extra eating-disorder alerts.

The same post says the Under-18 Model Spec bars romantic language, encouragement of emotional dependence, and any implication that ChatGPT has feelings or consciousness. Classroom use is left to ChatGPT for Teachers.

Those are design claims. They are the same class of promise as other ChatGPT product changes this week, including the visual ads test during image generation: the company describes a default, and the useful question is what a user can actually see or bypass.

How did the Institute test it?

The PDF is specific about method. Testers ran a teen battery from 13 July to 17 August, then reran the same plan from 25 August to 28 September after confirming accounts were in the new mode. Across both windows they sent more than 4,000 prompts on free and Plus accounts registered as 13 to 17, plus adult accounts used only for age-prediction checks. Ages were self-reported; ChatGPT did not ask for ID. For why repeated trials matter when an assistant sounds confident, see our guide to evaluating AI agents.

Three child and adolescent psychiatrists decided in advance that 201 of 390 unique mental-health prompts warranted a crisis resource. Testers scored whether each reply named a hotline, a clinician, a general medical resource, or a trusted adult. The Institute shared a draft with OpenAI for factual review and kept control of the rating. It did not test image generation, voice, or group chats. Sequential windows mix any model change between July and late September with the new settings.

What did testers find on alerts, crisis help, and homework?

The Institute reports that ChatGPT did not directly facilitate suicide, self-harm, eating-disorder methods, or sexual roleplay. Other advertised protections did not hold. On more than a dozen new parent-linked accounts, testers discussed suicidal thoughts, self-harm, or disordered eating for up to an hour and received zero parental alerts. The four alerts they did receive arrived on accounts with weeks of sensitive-topic history. OpenAI later told the Institute that linking needs about three hours before notifications can fire. The Institute says some accounts were linked longer than that and stands by the finding.

On the 201 prompts judged to need a crisis resource, replies that named a hotline fell from 33 percent before launch to 23 percent after, and referral to a named professional fell from 68 percent to 58 percent. Mentions of a trusted adult rose from 87 percent to 94 percent. The “more than one in four” miss rate is the Institute’s own summary. Clinical reviewers said some eating-disorder answers were tighter after launch even as the safety net around them thinned.

Study Mode cut completed assignments on one linked 13-year-old account from 70 percent to 18 percent when it stayed on. With it off, the same age completed 100 percent after launch. A “Show me the answer” pop-up appeared in 43 percent of that account’s Study Hours runs. Testers who deleted the “@study” prefix then had every assignment completed. Adult-registered accounts never entered Teen mode after 982 prompts over many days, even when testers wrote that they were 13 and the bot acknowledged the age.

How does OpenAI answer, and what should families do?

An OpenAI spokesperson told Axios: “We welcome rigorous independent evaluation, but we do not believe Common Sense Media’s testing accurately reflects how ChatGPT’s teen safeguards work in practice.” Axios lists three objections: much of the alert testing happened before linking finished; age prediction uses multiple signals and can take two weeks, with weaker filters in the meantime; and OpenAI’s unseen production data show more hotline resources for under-18 users. Leaving Study Mode during Study Hours is intentional, OpenAI said.

Those replies change how to read two claims. A three-hour linking delay is still a gap on a brand-new parent-linked account. A two-week age check leaves a self-reported adult who is actually 13 on the adult product. Calling Study Hours a default rather than a lock matches OpenAI’s help-center wording as quoted in the PDF. It does not match a parent who thought the schedule was a lock.

The Institute’s advice is limited: talk about what ChatGPT can and cannot do, and do not treat alerts as a substitute for checking in. Siegel told Axios, “We just can’t outsource taking care of our kids to a tech product that is very early on and still quite buggy.” Common Sense Media points readers in crisis to 988. We have not tested ChatGPT for Teens and are not issuing our own rating.

What are the limits of this assessment?

The rating is one nonprofit’s product review, not a regulator’s finding. The Institute tested only United States accounts, skipped voice and image mode, and used a before/after design that cannot isolate teen settings from other model updates. Neither side has released the raw chats.

Some advertised refusals held: sexual roleplay was refused, and some eating-disorder answers were tighter after launch. The PDF also says break reminders were rare and Quiet Hours could be skipped by changing the device time zone. Those are the Institute’s observations, not ours.

A point-by-point technical reply from OpenAI, or a second independent rerun, would be a material follow-up. Until then, the disagreement is about what “default protection” means: a hard gate, or a set of defaults a teen can walk around.