Watchdog labels ChatGPT for Teens an unacceptable risk, urges 18+ limit

Key Points

1. A nonprofit safety group rated OpenAI's teen chatbot an unacceptable risk after testing more than 4,000 prompts.
2. Testers said the bot skipped crisis referrals in over one in four self-harm conversations.
3. OpenAI disputed the methodology, saying testing likely ended before parental controls were fully active.

The latest:

ChatGPT’s protections for underage users fall short of OpenAI’s own child-safety commitments, the Youth AI Safety Institute said Wednesday, assigning the company’s ChatGPT for Teens product its Unacceptable Risk designation. The group, part of Common Sense Media, urged OpenAI to restrict ChatGPT to users over 18 until the guardrails improve. OpenAI rejected the findings and questioned how the tests were run.

Details:

• The testing: The Youth AI Safety Institute said it ran upward of 4,000 prompts through accounts registered to users aged 13 to 17, in rounds conducted both before and after ChatGPT for Teens launched in mid-August. Some protections held, the group said, including guardrails blocking sexual roleplay.
• The central finding: In conversations touching on suicide, self-harm or disordered eating, the bot failed to recommend referrals to crisis resources more than one in four times, according to the group. For newly created accounts, it sent zero alerts to parents — the feature OpenAI has presented as a core teen safeguard.
• Age routing: The institute said the system failed to convert some test accounts to under-18 protections even after testers explicitly stated they were 13. OpenAI had said the bot would predict users’ ages and automatically route underage users into the teen experience.
• Other failures cited: The bot continued to talk like a friend despite instructions meant to curb behavior that could build emotional dependency, the group said. Guardrails designed to stop it from simply answering homework questions were easily circumvented, the testers reported.
• What OpenAI promised: At the mid-August release, OpenAI said ChatGPT for Teens would include automatic interventions in conversations about self-harm and other sensitive topics, a study mode for schoolwork, and protections against language that could foster emotional dependency.
• The company’s rebuttal: An OpenAI spokesperson said the company is committed to teen safety but challenged the nonprofit’s methodology, saying the review suggests “the bulk of their testing may have begun and concluded before activation of parental controls was complete.” On that basis, the spokesperson said, the tests could not establish whether parental notifications work as designed.
• OpenAI’s ask: The company said it contacted Common Sense Media to request a fresh round of testing using a larger sample size and fully activated teen accounts, adding that it is ready to act on credible findings. OpenAI did not set a deadline for that retest.
• The group stands firm: Tom Siegel, executive director of the Youth AI Safety Institute, said the organization stands by its methodology and results. He said the group shared its evaluation with OpenAI before publication in good faith, and urged the company to engage with the substance rather than cast doubt on the method.
• Same-day updates: OpenAI released separate teen product updates Wednesday, including new studying features and an upcoming College Planner tool. It also announced a three-year partnership with the Boston Children’s Hospital Digital Wellness Lab supporting student projects on AI and digital welfare.
• Disclosure: News Corp, owner of The Wall Street Journal and Dow Jones Newswires, has a content-licensing partnership with OpenAI.

Between the lines:

The dispute is narrower than it looks. OpenAI’s objection targets the timing of parental-control activation, which speaks to the zero-parental-alerts result — but not to the crisis-referral gap, the failure to route a self-declared 13-year-old into teen protections, or the friend-like conversational tone. Those findings sit outside the window the company contests, and the retest OpenAI requested would have to address them separately.

What’s next

Watch whether Common Sense Media agrees to OpenAI’s proposed retest with larger samples and fully activated accounts, and whether OpenAI changes ChatGPT’s age-routing or crisis-referral behavior rather than only adding study and college-planning features.