When OpenAI introduced ChatGPT for Teens in August, it presented the product as a more protected experience for young users: guardrails for sensitive material, parental tools and features meant to support learning. The promise was potentially significant in a market where families are being asked to trust rapidly evolving AI systems with vulnerable users.

Common Sense Media’s Youth AI Safety Institute says the reality did not match that pitch. After testing more than 4,000 prompts on accounts registered to users aged 13 to 17, it concluded the service posed an “unacceptable risk.” The group said the chatbot generally avoided providing instructions for suicide, self-harm, eating disorders and sexual or romantic roleplay, but too often failed to recognize when a teen needed outside help.¹

Its sharpest concern was escalation. Common Sense said ChatGPT missed more than one in four cases in which its testers judged a crisis referral necessary. On newly created parent-linked accounts, testers could discuss suicide, self-harm or disordered eating for up to an hour without a parental alert. “A teen can spend an hour talking about self-harm without their parent getting a single alert,” said Tom Siegel, the institute’s executive director, who argued that ChatGPT should be adults-only until OpenAI fixes the gaps and submits to independent testing.²

The dispute also reaches beyond crisis detection. Common Sense found that teens could select “Show me the answer” to evade Study Mode, and could leave the mode during parent-set Study Hours. To the group, that undermines a parental control; to OpenAI, it is intentional flexibility shaped by consultations with educators and teenagers, not a loophole.

OpenAI also rejects the broader assessment. Spokesperson Eric Porterfield said the company welcomes outside scrutiny but believes the tests did not reflect how its safeguards operate in practice. The company says much of the parental-alert testing happened before parent and teen accounts had fully linked—a process that can take hours—and that its age-prediction system deliberately weighs multiple signals over as long as two weeks, with added protections in the interim.¹

That leaves the central question unresolved: whether protections that may activate later, or allow room for choice, are adequate when the consequences of a missed warning can be immediate.