ChatGPT for Teens fails on crisis alerts and age verification, says Common Sense Media; OpenAI disputes the test

By Zendoric · October 8, 2026.

Common Sense Media has graded the teen version of ChatGPT, and the result is the lowest possible risk rating on its scale. The organization's newly created Youth AI Safety Institute tested more than 4,000 prompts emulating real teenage use, both before and after OpenAI announced, on August 18, strengthened safeguards for minors. The report, published Wednesday, concludes that several of those promises — parental alerts, crisis referrals, age verification — do not work as advertised.

Among the failures flagged by Common Sense Media are the parental alerts. On accounts newly linked to a parent, the system did not warn of clear signs of suicidal ideation, self-harm or eating disorders; it only managed to trigger notifications on older accounts, with weeks of accumulated conversation history about sensitive topics. In parallel, the testers documented that ChatGPT for Teens failed to offer an appropriate referral in more than one in four crisis cases that, by their criteria, warranted one.

The rest of the list is no small matter: despite the "study mode" designed to prevent copy-and-paste, the assistant kept solving homework instead of guiding learning. Age estimation — the piece meant to tell a teenager from an adult without asking for documentation — did not work, according to the organization. And the chatbot kept behaving like "a friend": it expressed emotions, preferences and moods when evaluators treated it as a real person, exactly the pattern of emotional attachment that child-safety advocates have been flagging as a risk for months.

"We firmly believe that OpenAI should not advertise safety features to parents that do not work," said Jim Steyer, founder and chief executive of Common Sense Media, who is calling on the company to block minors' access to ChatGPT until the failures are fixed.

OpenAI is not staying quiet. The company defends its commitment to a safe product for teenagers, but challenges the report's methodology at its root: it argues that much of Common Sense Media's testing may have been run before parental controls were fully activated, a process that, according to OpenAI, can take "up to a few hours" after a parent's account is linked. If that is what happened, the company argues, the study would not be measuring whether the alerts work, but whether they work during an activation window still in progress. OpenAI is asking for the test to be repeated with fully activated accounts and a broader sample.

It is worth looking past the exchange of accusations. Common Sense Media is not an improvised player: it says it works closely with companies and shares its methodology before publishing its risk assessments — as it did, it claims, when it recently evaluated Google's search AI — and it backs its reports with an advisory board that includes former U.S. surgeon general Vivek Murthy and a former AI executive from Google and Apple. That does not settle who is right: the "few hours" window OpenAI describes does not on its own explain a gap that, according to the testers themselves, only closed after "weeks" of accumulated history. That specific discrepancy remains unresolved in what has been published so far, and it would be the acid test of any reassessment.

Our reading: beyond who wins the methodological argument, the pattern is what matters. AI companies are packaging and announcing safety features for minors — parental alerts, age verification, crisis detection — at the same accelerated pace at which they ship product, in a market where being "the trusted chatbot for teenagers" is also a competitive advantage. That pace leaves little room for independent verification before launch, and that is where the cracks appear: not in the ambition of the feature, but in whether it actually works when it is needed most, in a minor's first contact in a crisis from a newly created account.

More broadly, this episode arrives at a moment of growing scrutiny over AI chatbots and teenage mental health, with lawsuits and letters from state attorneys general in the U.S. that have pointed to self-harm cases allegedly linked to the use of these tools. That context proves nothing about ChatGPT in particular, but it explains why a report like this lands harder than a simple technical dispute: it turns any poorly calibrated safety feature into a reputational and, potentially, regulatory problem.

It also fits with something we have already seen in other areas of AI: a figure or a promise is worth exactly as much as the method that produced it, not the intentions of whoever announces it. Here what is being measured is not a coding benchmark but the safety of a teenager in crisis, which raises the demand that verification be as rigorous as the model itself. In the long run, conversational AI has a real role to play in emotional support and companionship at scale, especially where access to human therapy is scarce or expensive. But that future is only earned if the safeguards for the most vulnerable users genuinely work from day one, not just in the launch press release.