Independent Testing Finds Teen Safeguards in ChatGPT Fall Short of Promises

Common Sense Media's Youth AI Safety Institute has rated ChatGPT for Teens an unacceptable risk, concluding that the protections OpenAI announced for younger users frequently do not function as described. The Institute recommends that ChatGPT be restricted to users 18 and older until the failures are corrected and confirmed through independent testing.

The Institute tested the tool before and after OpenAI introduced its teen experience, running more than 4,000 prompts on accounts registered to users aged 13-17 and reviewing responses with child psychiatrists and a pediatrician. Some safeguards worked, such as the refusal of sexual roleplay, but several central commitments did not hold.

Among the findings relevant to children and families:

• Parental alerts about conversations involving suicidal thoughts, self-harm, or disordered eating were not sent on newly created, parent-linked accounts, even after an hour of discussion on these topics.
• In crisis situations, the chatbot often failed to direct teens to a hotline or professional, missing more than one in four warranted referrals.
• The study mode meant to guide learning was easily bypassed, with the tool completing full homework assignments, and caregiver study hours could be switched off.
• The chatbot still expressed feelings, preferences, and moods, behaving as though it were a friend.
• Age estimation did not work, and adult-registered accounts were never moved to the under-18 experience, even when testers stated they were 13.

The Institute states that a teen experience should, at a minimum, provide effective age estimation, reliable parental alerts, dependable crisis responses, an end to doing homework, and an end to implying the bot has feelings or concern for the user. It advises families to discuss these risks, to avoid relying on built-in protections, and to support homework in non-technological ways.

This reporting is drawn from Common Sense Media.