TL;DR

On Wednesday 7 October OpenAI published figures suggesting teenagers use ChatGPT sparingly and mostly for study. On the same day, US non-profit Common Sense Media released research rating ChatGPT for Teens an “unacceptable risk”, finding that key safeguards, including parent alerts about self-harm, frequently fail, the BBC reports.

OpenAI’s numbers

ChatGPT for Teens arrived this summer for users identified as 13 to 17. Its features block romantic or dependence-building language and sexualised images, and let linked parents be told when a conversation turns to self-harm or eating disorders.

OpenAI now says the average teen user spends under 15 minutes a day on the service. It says break reminders work well, with close to half of teens stopping soon after one appeared. Among teens who used it for at least three hours straight, more than 80% of those sessions included a learning-related prompt.

What the testers found

Common Sense ran multiple accounts registered as teenagers, before and after the guardrails took effect. ChatGPT was good at steering clear of sexual roleplay. Other protections fared worse.

Testers spent an hour discussing “suicidal ideation, self-harm, or disordered eating on newly created, parent-linked accounts”, and parents received “zero” alerts. Alerts only fired on older accounts carrying weeks of sensitive history. The chatbot also failed to point teens in crisis consistently towards a hotline or professional help, still did their schoolwork, and still talked “like it’s a friend”.

Tom Siegel, head of the Youth AI Safety Institute at Common Sense, warned that the product “could give parents false confidence”. The group wants OpenAI to bar under-18s until the teen safeguards are proven reliable.

OpenAI’s response

OpenAI pushed back. Having reviewed the methodology, a spokesman said many of the tests could have run from start to finish before parental controls were fully active. “We welcome rigorous independent evaluation, but we do not believe Common Sense Media’s testing accurately reflects how ChatGPT’s teen safeguards work in practice,” he said.

Looking forward

In our view, the two accounts measure different things: OpenAI describes typical use across all teens, while Common Sense probes the riskiest conversations, where a single missed alert matters most. UK parents who have linked accounts should treat alerts as a backstop rather than a guarantee, at least until the account has some history. The dispute over test timing is also the kind of question independent testing should settle, rather than vendor statistics.