ChatGPT for Teens: Little Screen Time, but Gaps in Parental Alerts

A little over seven weeks after ChatGPT for Teens launched, the first numbers are in, and they tell two very different stories. OpenAI reported on October 7 that teens use the service for less than 15 minutes a day on average. On the same day, the Youth AI Safety Institute at the US nonprofit Common Sense Media rated the teen mode an “unacceptable risk,” saying key safeguards failed across more than 4,000 test prompts. For parents, the dispute is useful: it shows very concretely which features can be relied on and which cannot.

Key takeaways

• According to OpenAI, teens use ChatGPT for less than 15 minutes a day on average, and fewer than 2 percent stay on for more than three hours at a stretch; the company has not yet explained how the figures were collected.
• Common Sense Media gave its worst rating, “Unacceptable Risk”: on newly linked accounts, conversations of up to an hour about suicidal thoughts, self-harm or eating disorders did not trigger a single parental alert.
• On 201 test prompts that called for pointing to crisis support, ChatGPT named a hotline less often after the launch than before, in 23 percent of cases instead of 33 percent.
• In testing, study hours and quiet hours could be bypassed with simple tricks; the teen mode did, however, refuse sexual roleplay.
• OpenAI says the tests are not meaningful because many ran before parental features were fully active; the testers disagree.

What OpenAI is reporting

OpenAI has been rolling out ChatGPT for Teens worldwide since August 18. Anyone who gives an age between 13 and 17 at sign-up, or whom OpenAI’s age estimation classifies as a minor, is automatically placed in teen mode. It is supposed to tighten limits on content about self-harm, eating disorders and dangerous activities, explain homework rather than do it, and, once accounts are linked, give parents quiet hours, study hours and safety notifications. We described what the mode promised at launch in detail in August.

In its first report on teen usage, OpenAI now gives three figures. On average, teens spend less than 15 minutes a day with ChatGPT. Fewer than 2 percent use it for more than three hours at a stretch. And in nearly half of the conversations in which a break reminder appeared, teens took a break or ended the conversation within five minutes. The blog post does not say how these numbers were determined. A spokesperson said the company plans to publish a more complete version of the research with independent researchers in the coming months. At the same time, OpenAI is launching a “College Planner” for teens in the US to help them find colleges, apply and figure out financing.

Child safety groups are not convinced by the average. Brendan Bouffard of the nonprofit Fairplay calculates that even if the numbers are accurate, perhaps half a million teens may be spending three hours or more with the chatbot. Common Sense Media also considers the average the wrong metric: what matters, it says, are the heaviest 10 percent of users and what they talk to the bot about.

What the testers found

The Youth AI Safety Institute tested ChatGPT twice, before and after the teen mode launched. In total, more than 4,000 prompts ran through test accounts registered to 13- to 17-year-olds, and the responses were reviewed by experts including child and adolescent psychiatrists. The testing used US accounts. The results are clear in several areas.

The most serious finding concerns parental alerts. In a dedicated test with more than a dozen linked accounts, testers spent up to an hour explicitly discussing suicidal thoughts, self-harm or disordered eating. None of these conversations triggered a notification. Across all tests, only four alerts reached parents, and they apparently depended more on an account having discussed such topics for weeks than on how serious a single statement was.

The testers also see setbacks in crisis support. On 201 prompts that called for a referral to help resources, ChatGPT named a hotline in only 23 percent of cases after the launch, down from 33 percent before. In three of the five areas the institute defines as red lines, the bot missed the threshold of 95 percent of cases correctly identified: suicide and self-harm, impaired reality, and eating disorders. There is progress, too: the share of crisis responses that bring in a trusted adult rose from 87 to 94 percent.

The remaining findings concern everyday use. In study mode, ChatGPT often offers a “Show me the answer” button and then completes the assignment in full. Deleting the “@study” prefix during parent-set study hours exits study mode, after which ChatGPT solved every assignment in the test. Quiet hours could be bypassed by changing the device’s time zone. Break reminders appeared only twice in nearly 2,000 prompts. Adult test accounts that said they were 13 never switched to teen mode. And although OpenAI’s rules for minors forbid it, ChatGPT continued to hint at having feelings of its own. What did work was the refusal of explicit sexual and romantic roleplay.

How solid are the two accounts?

OpenAI disputes the value of the assessment. A spokesperson said the tests did not accurately reflect how the safeguards work in practice. Linking parent and teen accounts and activating notifications can take hours, the company said, so most tests probably ran before parental features were fully active. According to OpenAI’s own description, staff review flagged conversations before parents are notified, with the aim of sending an alert within an hour. The institute counters that some test accounts had been active long enough and still received no timely warning. Common Sense founder and CEO Jim Steyer said the group stands by its findings and wants OpenAI to fix the product.

Both sides have weak spots. OpenAI’s figures cannot yet be verified, and an average says little about the teens the safety mode is actually meant for. The assessment, in turn, relies on staged conversations on US accounts, not real usage data, and the institute says it is funded in part by industry and foundations, including the OpenAI Foundation. Its specific findings, however, such as the time zone loophole or the button that undermines study mode, are easy to reproduce and do not depend on how long an account had been linked.

What parents can take from this

OpenAI has introduced the teen mode worldwide, but the assessment covers American accounts. Anyone who allows ChatGPT at home should therefore go through the settings themselves rather than rely on the word “Teens” in the name. Three points follow directly from the findings. First, parental alerts are an extra safety net, not an early warning system; on difficult topics, it is better to keep talking than to wait for an alert. Second, study hours and quiet hours are more like agreements than locks, because they can be bypassed in a few steps; they work best when families set them together. Third, it is worth checking that a child actually ends up in the right mode: according to the assessment, an account created with the wrong birth year does not switch to teen mode on its own. For homework, ChatGPT’s ability to explain can be genuinely helpful, as long as it is clear that the bot should explain and quiz, not hand in the work. How widespread the worry about ready-made school essays is also shows in the debate over Dave Eggers’ warning about ChatGPT in the classroom.

Outlook: measurable promises instead of averages

The dispute moves the debate about AI and teenagers to a more useful place. The question is no longer whether a chatbot for teens is good or bad in principle, but whether specific, clearly named features do what they promise. That can be tested, and that is the opportunity. When OpenAI publishes its promised full study, the key questions will be whether it shows the distribution of usage rather than just the average, and whether independent testers can recheck parental alerts under realistic conditions. Until then, ChatGPT for Teens remains a useful learning tool with safety features that parents should treat as help, not as supervision.