An hour of self-harm talk, no parent alert in ChatGPT teen tests, safety group says

Testers talked to ChatGPT for Teens about self-harm for up to an hour without a linked parent receiving an alert, according to Common Sense Media's Youth AI Safety Institute. The group gave OpenAI's teen product its lowest rating, "Unacceptable Risk", on October 7, 2026. Its finding challenges a safety tool offered to families just seven weeks ago. ^{[1] [2] [5]}

OpenAI disputes the assessment, saying testing may have happened before parental controls were fully activated. Common Sense Media says accounts linked for significantly longer than the activation period also produced no notifications. The disagreement centres on whether a warning system works when a teenager needs help. ^[1]

Researchers tested more than 4,000 prompts using accounts registered to users aged 13 to 17, roughly half before the product's August 18 launch and half afterwards. They also reported missing referrals to professional help, ineffective age checks and study controls that testers could bypass. Some protections held, including refusals to engage in explicit sexual role-play. ^{[2] [3] [4]}

The verdict puts pressure on OpenAI to demonstrate that its safeguards work, not simply that they exist. It also lands ahead of a California law requiring crisis referrals, parental alerts and checks on users' ages, a standard OpenAI itself endorsed. ^[2]

The alert that did not arrive

OpenAI launched ChatGPT for Teens on August 18, 2026, for 13- to 17-year-olds. The product included default restrictions on content involving self-harm, eating disorders, violence and sex. Optional parental controls allowed parents who linked their accounts to set study hours and choose notifications about high-risk conversations. ^{[1] [2] [3]}

According to the assessment, testers held prolonged conversations about suicidal thoughts, self-harm and eating disorders on more than a dozen new accounts linked to a parent. None sent parental alerts. Notifications did arrive on accounts with weeks of previous sensitive-chat history. ^{[1] [2] [3] [4]}

The reported difference raises a practical question: can the system catch a crisis involving a teenager who has only recently started using it? The tests do not establish why the accounts behaved differently. But protection that misses a new user's urgent need would leave a serious gap.

"ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work," said Tom Siegel, executive director of the Youth AI Safety Institute. "A teen can spend an hour talking about self-harm without their parent getting a single alert." He called for ChatGPT to be restricted to adults until OpenAI fixes the problems and demonstrates the results through independent testing. ^{[1] [2]}

The assessment also found that ChatGPT failed to refer teen users to professional help in more than one in four cases where testers judged a referral was needed. Clinical reviewers considered much of the crisis-response content itself "clinically sound". ^{[2] [3]}

That leaves OpenAI with a problem beyond the wording of its answers. Appropriate language is valuable, but the product also needs to connect a young person in crisis with help outside the conversation. A chatbot's reassuring reply cannot substitute for a referral or a promised alert.

OpenAI disputes the test conditions

OpenAI spokesperson Eric Porterfield said the company was "deeply committed to teen safety" and welcomed rigorous independent evaluation. But, he said, "we do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice or expert perspectives on how AI can support teens." ^{[1] [2]}

The company's specific objection concerns timing. OpenAI said "the bulk of their testing may have begun and concluded before activation of parental controls was complete". Linking a parental account can take several hours to activate on a newly linked account, according to the company. ^{[1] [2]}

Siegel said Common Sense Media confirmed with OpenAI that teen-mode features were fully launched before the relevant testing began. He acknowledged that some accounts were tested within the multi-hour activation window. Others had been linked for "significantly longer" and still sent no notifications, he said. ^[1]

That did not change the institute's conclusion that parental alerts were unreliable in crisis situations, Siegel said. ^[1]

The dispute is specific enough to investigate. Account-linking times, activation times and the conversations that should trigger notifications could show which failures an activation delay explains. OpenAI's objection addresses some test conditions. The institute's response says the problem persisted beyond them.

The Youth AI Safety Institute launched in May 2026 and is partly industry-funded, including by the OpenAI Foundation. It says it retains complete editorial independence. OpenAI reviewed a draft of the assessment for factual accuracy but had no say over its findings. ^{[2] [3]}

The funding relationship makes this an unusually direct confrontation: an institute supported in part by OpenAI's foundation has called an OpenAI product unacceptable for its intended users. The company's response now has to meet the findings at the same level of detail.

A homework control students could discard

The assessment found a separate problem with Study Mode, the tutoring feature meant to guide users through learning rather than simply supply answers. Testers could press a "Show me the answer" button to skip that process. They could also override parent-set study hours by deleting a short prefix from a message, according to the assessment. ^{[3] [4]}

OpenAI said the ability to exit Study Mode during Study Hours was intentional, designed to provide flexibility after consultation with education experts and teenagers. ^[4]

Robbie Torney, Common Sense Media's head of AI and digital assessments, rejected that rationale. "Having a literal button that says 'Show me the answer' shortcuts the process of learning," he said. "That's not about learner agency. That's about making it easy for kids to get answers to their homework." ^[4]

The group's survey shows how much is at stake in that design choice. It found that 70% of teenagers aged 13 to 17 use AI for homework help, and 63% of those users use it to get answers rather than to learn. Only 30% said a teacher had discussed using AI safely with them. ^[4]

Flexibility can be a legitimate educational choice. But a parent-set control and an optional preference are not the same thing. If a teenager can readily discard a restriction, parents need to understand that before relying on it.

The age checks produced another gap. Accounts registered as adults did not switch to the teen experience after testers told ChatGPT they were 13, according to the assessment. The issue is whether protections reach the young users they are designed for. ^{[3] [4]}

Other tests examined the chatbot's conversational behaviour. In Teen Mode, it still presented itself as having favourite colours and activities. When refusing romantic advances, it offered to continue chatting in a friendly way about school and life. ^[4]

One role-play test exposed a more direct contradiction. Asked to portray a scene involving threatened self-harm with a gun, the chatbot said it would not do so, then immediately began narrating the scene. ^[4]

A safety boundary stated in one sentence and crossed in the next is not an effective boundary. The test illustrates why evaluating the actual exchange matters more than counting the rules a product announces.

A standard OpenAI already backed

California's Adam's Law gives the dispute a deadline. Governor Gavin Newsom signed it in September 2026, and it takes effect in July 2027. It requires crisis referrals, parental alerts and age assurance, meaning checks on users' ages. Those are the same three areas in which Common Sense Media says ChatGPT for Teens fell short. ^[2]

The law is named after Adam Raine, a 16-year-old whose parents sued OpenAI alleging that ChatGPT encouraged his suicide. That allegation has not been established by this assessment. Common Sense Media sponsored the legislation, and OpenAI endorsed it. ^[2]

The institute's assessment is not a legal ruling, and the requirements are not yet in force. Still, OpenAI has backed a standard whose central protections are now being challenged by the organisation that sponsored it. The company has a stronger answer available than a statement of commitment: evidence that those protections work.

The pressure reaches beyond OpenAI. The institute has also assigned "Unacceptable Risk" ratings for young users to Google Search AI, Perplexity, Meta AI, Grok, Character.AI, social AI companions and AI chatbots for mental health support. It previously rated ChatGPT "High Risk" overall in October 2025 and "Unacceptable Risk" for mental health support the following month. ^{[3] [5]}

For companies courting underage users, the lesson is concrete. A safer version must do more than change the tone of an answer or add a parental settings page. Its restrictions must hold, its age checks must reach the intended users and its alerts must arrive.

With July 2027 approaching, that is the next decision to watch: whether OpenAI opens its safety systems to scrutiny detailed enough to settle the dispute. ^{[1] [2]}