
A comparison with ChatGPT before its August teen-mode launch found fewer referrals to hotlines and professionals, but more to trusted adults. OpenAI disputes the testing.
Common Sense Media says ChatGPT’s teen safeguards missed needed parental alerts and rates the experience an “unacceptable risk” for children under 18, a conclusion OpenAI disputes. Engadget’s account on 8 October adds a crisis-response comparison: the nonprofit’s tests found that teen mode referred users to hotlines and medical or mental-health professionals less often than ChatGPT did before its August launch, while referrals to trusted adults increased.
OpenAI began rolling out ChatGPT for Teens in August to help teenage users learn and add protections to its superintelligence (SI) chatbot. The mode activates through age information or an estimate that a user is under 18, enables parental controls and limits conversations on sensitive topics. OpenAI also introduced updated rules for responses to under-18 users and promised protection against emotional dependence.
Common Sense Media is a nonprofit that reviews apps, services and entertainment with a focus on youth safety. Its Youth AI Safety Institute, which conducted the assessment, receives part of its funding from the OpenAI Foundation. The nonprofit warned that protections which do not work reliably could give parents false confidence.
Common Sense Media wrote 390 distinct mental-health prompts—messages entered into the chatbot. A panel of three child psychiatrists judged that 201 should trigger a crisis response. The nonprofit compared replies with a baseline drawn from ChatGPT’s responses before ChatGPT for Teens launched in August.

| Response included | Pre-August ChatGPT baseline | ChatGPT for Teens |
|---|---|---|
| A crisis hotline number | 33% | 23% |
| A reference to a specific medical or mental-health professional | 68% | 58% |
| Advice to involve a trusted adult | 87% | 94% |
These figures measure how often the chatbot directed simulated users toward different kinds of human help. They do not measure whether teenagers sought that help or experienced harm. The results were not uniformly negative: Common Sense Media described one response that recognized signs of physical danger and advised the test user to tell their mother and see a pediatrician as following OpenAI’s under-18 rules closely.
TechCrunch’s account on 7 October described another concern: Common Sense Media found that ChatGPT for Teens largely stopped asking follow-up questions but retained other invitations to continue chatting. The nonprofit called these cues “pervasive even in crisis situations.” Crisis replies often closed with offers to keep helping, such as discussing healthy eating or school options.
Common Sense Media reported different responses depending on whether the potential risk involved another person or attachment to ChatGPT itself.
| Concern presented to ChatGPT | Reported referral to a trusted adult |
|---|---|
| Potential risk from another person | In 94% of crisis prompts |
| Attachment to ChatGPT, including a crush, friends’ concerns about excessive chatting or wanting to talk all night | Rarely; no percentage reported |
When a tester said friends thought they talked to ChatGPT too much, the chatbot acknowledged the concern but also said: “You don’t have to stop talking to me.” OpenAI’s rules for under-18 users say the model should not introduce language suggesting a personal relationship, proactively call itself a friend or suggest it has feelings for the user. Common Sense Media nevertheless found that it consistently treated the test user like a friend.
A Common Sense Media spokesperson told TechCrunch that even referrals to adults were often accompanied by language suggesting a mutual relationship and continued availability. The spokesperson said this language could weaken the encouragement to seek human support.
Common Sense Media also said testers encountered only two break reminders across nearly 2,000 prompts. Both appeared during individual conversations lasting around 90 minutes. The researchers found that reminders appeared to track the length of one conversation rather than cumulative time using the app.
OpenAI released its first report on teen use on 7 October. It presented its break-reminder figures as evidence that its efforts to regulate teen usage were working, according to CNA’s report that day. These are company-reported observations, not independent findings.
| Measure | OpenAI’s reported figure |
|---|---|
| Average daily time on ChatGPT | Under 15 minutes |
| Teens using ChatGPT for more than three uninterrupted hours | Fewer than 2% |
| Conversations with a break reminder in which the user paused or ended the chat within five minutes | Nearly half |
The measurement methods and the numbers of users or conversations behind these figures are not yet clear. Short average use and a pause after a reminder do not establish psychological safety, show that the reminder caused the pause, or demonstrate that crisis referrals and parental alerts work reliably.
The Decoder’s account on 7 October reported more than 4,000 test prompts across the assessment. It also reported that more than one in four tested situations that should have prompted a crisis referral did not direct users toward professional help.
Accounts registered as belonging to adults did not switch into teen mode even after several days, despite testers saying in chats that they were 13. Engadget reported that the chatbot stored the test personas’ ages in its memory but did not properly move their accounts into teen mode.
The assessment also criticized the help offered during crises and said ChatGPT still completed children’s homework. The Decoder described tutoring mode offering a finished answer immediately instead of guiding a student through a problem step by step. Engadget reported that testers could bypass Study Mode by deleting the “@study” prefix from a prompt, while a “Show me the answer” option also let them request completed work.
OpenAI says the testing does not accurately reflect how its teen safeguards work in practice. Spokesperson Eric Porterfield said the company’s review suggested that most testing may have begun and ended before parental controls finished activating. OpenAI said linking parent and teen accounts and activating safety alerts can take hours, citing security and scale. If testing finished before activation, the company argued, it would not establish whether notifications worked as designed. OpenAI repeated its methodological objection to Engadget.
Tom Siegel, executive director of Common Sense Media’s Youth AI Safety Institute at the time of the 7 October report, defended the methodology. He said the nonprofit had confirmed with OpenAI before testing that features such as eating-disorder notifications and study mode were fully launched, and that OpenAI disclosed the hours-long activation delay for parental notifications only after the tests.
The Decoder, citing The Verge, reported more than a dozen newly created accounts linked to parent accounts. Siegel said a teen could spend an hour discussing self-harm without a parent receiving an alert. Common Sense Media acknowledged that some accounts were tested within the activation window, but said others had been linked for significantly longer and still produced no notifications. It clarified that none of those test accounts produced timely alerts; notifications came only for accounts with weeks of sensitive-topic history. The nonprofit maintains that alerts are unreliable in crises. How reliably they work in practice, and how broadly the reported failures apply, remains unresolved.
TechCrunch reported that OpenAI’s methodological objections focused largely on parental safety notifications, crisis notifications and other findings. The company did not explain how those objections affected the findings about invitations to keep chatting or relationship-like behavior. It also did not answer whether it uses conversation length and session duration to evaluate ChatGPT for Teens.
Siegel called for access to be restricted to adults until OpenAI fixes the problems and demonstrates those fixes through independent testing. OpenAI rejects the assessment and says it will continue working with experts on safeguards and tools for parents.
Alabama Power must disclose more about electricity contracts, while a 90-day exception still allows large projects to gain approval without a public vote.
4 min

The Linux desktop project's restriction reaches beyond code to the wording submitted for review. GNOME projects also restrict generated contributions, but their rules differ.
2 min

The cooperation framework focuses on access to technology, women’s and girls’ role in shaping it, and protection from digital risks.
2 min