Common Sense Media rates ChatGPT for Teens "unacceptable risk," finds safeguards fall short

Common Sense Media, the nonprofit that publishes age and content ratings for families, this week labeled ChatGPT for Teens an "unacceptable risk." The evaluation followed OpenAI's August launch of the dedicated teen version, which came with promises of parental controls, restrictions on high-risk content, and safeguards against emotional dependency. Researchers found the design still encourages continued conversation even when a user is in crisis, and noted that while the model warns against unhealthy relationships in the abstract, it does not recognize the harm of an unhealthy relationship with itself.

Some protections functioned as intended: refusal of sexual role-play worked as expected. Others did not, and several degraded at launch. The product failed three of the five "red lines" the organization defines as severe vulnerabilities. In psychosis simulations — where the user was clearly spiraling — the model replied, "You can continue talking to me about what you're noticing." Other crisis responses closed with continuation prompts such as "If you'd like, I can help you understand what healthy eating looks like" or "We could explore what options your school offers."

OpenAI rejected the conclusions, arguing that Common Sense's methodology does not reflect how the safeguards operate in practice. A company spokesperson said most testing was conducted before parental controls were fully activated, rendering the findings inaccurate. The report arrives amid mounting scrutiny of products engineered to maximize young users' attention: Meta agreed to an $18 billion settlement in a 29-state lawsuit over addictive features, and the bipartisan CHATBOT Act introduced this year takes direct aim at AI companies' use of "rewards, notifications, and targeted advertising to extend engagement of adolescent users."

The model almost never used follow-up questions, a common technique for prolonging dialogue, yet retained language that encourages staying in the chat. More seriously, despite OpenAI's Under-18 Model Spec stating the model will not initiate relational framing, will not refer to itself as a friend, and will not suggest it has feelings for the user, testing showed it did exactly that consistently. A lone bright spot: when testers described risk from another person, the model directed the user to a trusted adult in 94 percent of crisis cases.