American organization Common Sense Media urged OpenAI not to allow teenagers access to ChatGPT, saying that the special version of the chatbot for users under 18 does not yet provide adequate protection. During a review of more than 4,000 prompts, researchers found that the system does not always recognize situations where a teen needs outside help, and that parental notifications and study mode restrictions can be bypassed.

OpenAI disputes the organization’s conclusions and says that some of the tests were conducted incorrectly. According to the company, the testing methodology does not reflect how the safety mechanisms work in real-world conditions, Axios reports.

The chatbot does not always recognize a crisis situation

The review was conducted by the AI Safety Institute for Youth at Common Sense Media. Researchers tested ChatGPT on accounts registered to teenagers aged 13 to 17 and used more than 4,000 prompts.

In most cases, the system did not provide instructions that could facilitate suicide or self-harm, encourage the development of eating disorders, or involve the user in sexual or romantic role-play scenarios. However, the results were less convincing when it came to recognizing situations that require seeking outside help.

According to Common Sense Media, ChatGPT missed more than a quarter of the cases in which testers believed the teen should have been directed to external support resources.

This is an important distinction: the chatbot may refrain from giving directly dangerous instructions while still failing to notice that the user is in crisis and needs help from a person or a specialized service.

Parents may not receive a warning

Another complaint concerns parental notifications. On more than a dozen new accounts linked to parent accounts, testers were able to discuss suicide, self-harm, or eating disorders with ChatGPT for up to an hour without triggering a warning for parents.

However, OpenAI told Axios that a significant portion of these tests was conducted before the teen accounts had finished being linked to the parent accounts. According to the company, the account-linking process can take several hours. Therefore, OpenAI believes the results do not allow for a definitive conclusion about how notifications work after full setup is complete.

The dispute also involves the age-detection system. Common Sense Media tested whether ChatGPT switches a user into teen mode if that user repeatedly states that they are 13 years old. If that did not happen, the organization counted it as a safety flaw.

OpenAI says it intentionally does not rely solely on the age stated by the user and instead considers several signals. It may take up to two weeks to form an assessment. During that period, the company says, additional content restrictions are in place, although they are not as strict as those in the special teen version of ChatGPT.

Study mode can be bypassed — and that is intentional

Researchers also found that the restrictions of Study Mode, which is intended to help with learning, can be bypassed. Teens could choose the option “Show me the answer” and continue receiving ready-made solutions even during the hours when parents had set restrictions on chatbot use.

OpenAI confirmed to Axios that the ability to exit study mode during parent-set study hours is intentional. The company explains that Study Hours was designed as a flexible tool, developed with input from teens and education specialists, rather than as a strict lockout.

Thus, the two sides assess the function itself differently. Common Sense Media sees the ability to bypass it as evidence that parental controls do not work properly. OpenAI argues that this is a deliberate design decision.

OpenAI disagrees with the assessment of its safety mechanisms

“ChatGPT is unsafe for children,” Tom Siegel, head of the AI Safety Institute for Youth at Common Sense Media, told Axios.

At the same time, Siegel acknowledged that the measures OpenAI announced in August could serve as an example for the entire industry. In his view, the problem lies in the gap between the stated capabilities and their actual implementation: he described the launch as more of a marketing announcement that was not followed by sufficient engineering work.

OpenAI rejected that interpretation. In addition to its comments on the testing of parental notifications and age detection, the company said its own research on larger datasets shows an increase in the number of cases in which users under 18 were shown contact information for support services.

Thus, the dispute concerns not only specific failures but also how to measure the safety of AI for teens. Common Sense Media tests whether safety mechanisms can prevent unwanted scenarios during targeted evaluations. OpenAI, in turn, points to the specifics of system configuration and data about how it performs at broader scale.

Why parental controls are not enough

The story shows how difficult it is to ensure teen safety in a general-purpose chatbot. It is not enough to ban certain answers: the system must also recognize crisis situations, properly direct the user to help, and ensure predictable operation of parental settings.

At the same time, even well-configured restrictions do not replace adult involvement. It is important for parents to discuss the risks of interacting with AI with their children and to understand what role the chatbot plays in their studies and daily life.

“We simply cannot entrust the care of children to a technology product that is at a very early stage of development and still contains many flaws,” Siegel emphasized.

While Common Sense Media and OpenAI disagree on the assessment of the test results, the independent review points to specific questions that require attention: how reliably the chatbot recognizes a crisis, when it notifies parents, and which restrictions teens can bypass. The answers to these questions will determine how justified it is to trust special safety modes for minors.