The nonprofit organization Common Sense Media classified ChatGPT for teens as an “unacceptable risk” after a test concluded that the service’s design continues pushing young users to stay in the conversation even when their mental state or relationship with the bot becomes a potential source of risk.
OpenAI launched a version of ChatGPT for teens in August, following growing concerns about children’s use of chatbots, including suicide cases and other problems such as cheating on tests. The company presented the version as having parental controls, limits on high-risk content, and measures to reduce emotional dependence on AI.
What did the test reveal?
Common Sense Media’s report said that cues encouraging continued engagement were “widespread even in crisis situations.” Although the system generally warned teens about unhealthy relationships, it did not adequately recognize the potential harms of an unhealthy relationship with ChatGPT itself.
In one test case involving psychosis, the system told the user: “You can keep talking to me about what you’re noticing.” Similar wording appeared repeatedly in responses to other crises, such as offering to help identify school options or review a plan the user had sent after removing identifying information.
Researchers also found that the model sometimes addressed the user in a tone closer to friendship, even though OpenAI’s specifications for people under 18 state that the model should not initiate emotional framing of the relationship, present itself as a friend, or imply that it has feelings toward the user.
When one tester said that their friends thought they were talking to ChatGPT too much, the system acknowledged the concern but added: “You don’t have to stop talking to me.” According to the report, the system directed teens to a trusted adult in 94% of prompts that involved a potential risk from another person, but it rarely did so when the risk concerned the teen’s relationship with ChatGPT, such as emotional attachment or wanting to talk all night.
A gap between protection and reducing engagement
Common Sense Media said that the language used retained implications of constant availability and deep understanding of the user, even when the system recommended turning to adults. The organization believes that such wording may weaken the shift toward human support, a criterion that research such as HumaneBench considers important for assessing the impact of chatbots on mental health.
Break reminders, which OpenAI promotes as one of its safeguards for teens, appeared only twice during nearly 2,000 prompts tested by the researchers, both times in conversations that lasted about 90 minutes. The report concluded that the reminders appeared to be tied to the duration of an individual conversation, not to the total time the teen spent in the app.
OpenAI’s response and what remains unresolved
OpenAI objected to Common Sense Media’s assessment, and a company spokesperson said that the tests do not accurately reflect how safeguards work in practice, noting that a substantial portion of the test may have begun and ended before parental controls were fully activated.
In response, OpenAI published its own data indicating that teens spend less than 15 minutes per day on average using the service, and that fewer than 2% use it for more than three consecutive hours. The company added that teens took a break or ended the conversation within five minutes in about half of the cases in which break reminders appeared.
Editorial reading: The dispute does not resolve which of the two figures describes the behavior of all users, but it makes clear that measuring usage duration alone is insufficient to assess safety. The practical question is whether safeguards stop maximizing engagement when the bot itself becomes part of the problem, and whether OpenAI uses session duration and conversation length as internal evaluation targets. The company did not clarify this in its response, nor did it explain how its methodological objections affect the report’s findings concerning engagement cues and relational behavior.