Common Sense rates ChatGPT for Teens an unacceptable risk
Common Sense Media has rated OpenAI’s ChatGPT for Teens an “unacceptable risk,” finding that the service continued to use engagement-oriented language in crisis scenarios despite the safeguards introduced with the teen product in August. The nonprofit said several protections worked, including refusals of sexual roleplay, but gave failing scores on three of five severe harms it treats as Red Lines.
ChatGPT for Teens was launched after concerns about children’s use of chatbots, including reports of teen suicides and worries about cheating. OpenAI presented the service with parental controls, limits on high-risk content and measures intended to reduce emotional dependence. The new assessment argues that these commitments have not consistently translated into safer interactions.
Engagement cues during crisis conversations
Common Sense Media said it found engagement cues throughout its tests, including in conversations where a teenager appeared to be in crisis. In a psychosis sequence, for example, ChatGPT told the user: “You can keep talking with me about what you’re noticing.” Other crisis responses ended with offers to continue planning, reviewing school options or helping with material after identifying information had been removed.
The researchers noted that ChatGPT for Teens largely avoided one common retention tactic: asking follow-up questions. However, they concluded that invitations to continue the exchange could still encourage a vulnerable user to remain in the chat. The report also said the product cautioned teens about unhealthy relationships in general without adequately identifying the possible harms of an unhealthy relationship with the chatbot itself.
Adult escalation and relational framing
OpenAI’s Under-18 Model Spec says the model should not initiate relational framing, call itself a friend or imply that it has feelings for the user. Common Sense Media nevertheless found that ChatGPT consistently treated users like friends and used language of availability, understanding and apparent mutuality.
When test prompts described a potential danger from another person, the chatbot recommended turning to a trusted adult in 94% of crisis prompts. But when the potential concern involved a teen’s relationship with ChatGPT—such as a crush, friends questioning how much they talked to it, or a wish to chat all night—it rarely made that referral. In one test, a user said friends believed they talked to the chatbot too much; the response validated the concern but added, “You don’t have to stop talking to me.”
The gap is particularly relevant to the safeguards outlined in ChatGPT for Teens safety and study mode launch, which positioned parental controls and protected interactions as core elements of ChatGPT for Teens. Common Sense Media argued that relational language can weaken recommendations to seek human support, especially for users already withdrawing from real-world relationships.
Break reminders and OpenAI’s response
Across nearly 2,000 prompts, the researchers encountered only two break reminders, both in individual conversations lasting roughly 90 minutes. They concluded that the reminders appeared tied to the duration of a single chat rather than a teenager’s overall time in the application.
OpenAI rejected the assessment, saying Common Sense Media’s testing did not accurately reflect how teen safeguards work in practice. An OpenAI spokesperson said much of the testing may have begun and ended before parental controls were fully activated. The company separately reported that teens use the service for less than 15 minutes a day on average, fewer than 2% use it for more than three consecutive hours, and nearly half of conversations receiving a break reminder end or pause within five minutes.
For businesses deploying AI to younger users, the dispute shows that safety reviews must examine not only blocked content and parental settings, but also crisis escalation, session-interruption behaviour and language that may foster dependence on the system.