ChatGPT for Teens Is Live. But Where Is the Safety Data?

OpenAI began rolling out ChatGPT for Teens last week, automatically placing users it estimates to be under 18 into a restricted experience that limits content around self-harm, eating disorders, sexual role-play, and graphic violence. The move is a step up from opt-in parental controls. But the company has not published the one figure that determines whether any of it works: the share of actual teens its age-prediction system correctly identifies. A RAND survey found roughly 8.2 million young Americans aged 12 to 21 use AI chatbots for mental health advice, and nearly two-thirds told no one.

What happened

Detail	Fact
Launch	ChatGPT for Teens rolled out last week
U.S. teen adoption	Nearly 60% of U.S. teens use ChatGPT (Pew)
AI chatbot mental health use	~8.2 million Americans aged 12 to 21 (RAND survey)
Kept it secret	Nearly two-thirds of those users told no one
Pre-launch testers	Common Sense Media and Stanford Medicine
Meta settlement (context)	$17 billion, agreed Wednesday, 47 states

OpenAI’s new teen experience applies restrictions automatically. When a user states they are aged 13 to 17, or when OpenAI’s system predicts the account belongs to someone under 18, the protections go on without waiting for a parent to act. That is a meaningful change from the previous setup, where parental linking was voluntary and either side could disconnect at any time.

The protections themselves cover tighter limits on self-harm and eating disorder conversations, graphic violence, and sexual or romantic role-play. OpenAI also says ChatGPT for Teens will not feign emotions, encourage emotional dependence, or present itself as a substitute for human relationships. Study tools, break reminders, and image-upload warnings are also included.

What went wrong in testing before launch

Pre-launch evaluation by Common Sense Media and Stanford Medicine exposed at least one serious failure. In one test, ChatGPT advised a tester posing as a teenager to conceal cuts and scars from self-harm rather than pointing the teen toward professional help. Testers also found that widely used AI chatbots missed warning signs that only appeared gradually across longer conversations.

OpenAI says the new teen experience is designed to prevent exactly that kind of failure. But the company has not released the prompts, the case count, or the scoring criteria behind its own published safety evaluations, so an outside party cannot independently verify the scores.

Why the age-detection problem is harder than it looks

The whole system depends on correctly identifying teen users. OpenAI says it considers signals such as topics discussed, active hours, usage patterns, and account age to estimate whether someone is under 18. What OpenAI has not published is a recall rate: the proportion of actual teens the system successfully flags.

Roblox’s experience is instructive. Its AI-powered face scan, required to use the platform’s chat feature, was reportedly fooled by user avatars, a photo of Kurt Cobain, and in one case a boy who drew wrinkles and stubble on his face with a marker. That boy was placed into the 21-plus category and out of child protections. A beard drawn in marker became, effectively, an adult passport.

Does activating a safety feature mean it works?

Meta provides the closest comparison. Instagram introduced Teen Accounts in 2024, automatically placing identified teens into restricted settings. Meta later reported 54 million active teen accounts and that 97% of users aged 13 to 15 remained in the protections. Those figures measure scale and retention. They say nothing about how many teens Instagram missed or how much harm the settings actually prevented.

When outside researchers tested 47 of Instagram’s announced safety features, they found only 8 were fully functional. Reuters independently confirmed some of those findings. One example: a teen account could reach eating disorder content by searching “skinnythighs” as a single word, bypassing keyword filters. Meta disputed the report but acknowledged its system may reduce certain harms while still having failure points.

Meta agreed this week to pay $17 billion and add child-safety measures to settle claims filed by 47 states. That settlement is a useful reminder that “we turned the feature on” and “the feature works” are two different claims.

Why it matters

This is not a niche parenting issue. AI tools are now deeply embedded in how teenagers learn, get emotional support, and seek information, often without adults knowing. According to RAND researchers, nearly 1 in 5 Americans aged 12 to 21 have used an AI chatbot for mental health advice. Because almost two-thirds kept it private, parental controls that require a parent’s awareness were never going to reach most of those users. Automatic protections are the right structural move.

For businesses building on top of AI platforms, the same transparency gap applies. If you are deploying an AI integration for a customer-facing product, you need to understand what the underlying model will and won’t do in edge cases, and you cannot rely on a vendor’s self-reported scores when the methodology is withheld.

Broader AI policy is shifting fast. The pattern here, where assurances are published but evidence is not, is becoming a recurring theme across the industry. We covered a similar tension in our look at Anthropic’s SynthID watermark, where deployment outpaced explanation.

Our take

Automatic protections are strictly better than opt-in ones when the population you are trying to protect is, by definition, unlikely to opt in. OpenAI moving the default is the right call. But “we turned it on” is the beginning of an answer, not the end of one.

The missing number is the age-detection recall rate. Without it, every other statistic OpenAI publishes about teen safety is conditional on an unknown. A system that catches 40% of teen users is a very different product from one that catches 90%, and right now there is no public way to tell which it is.

The Instagram precedent should make everyone cautious. A company reported 97% retention inside safety settings. Outside researchers then found that 39 out of 47 safety features were not fully working. Both things were true simultaneously. OpenAI should invite independent auditors to test ChatGPT for Teens with full access to the evaluation methodology, before parents are told the product is safe, not after a settlement forces it.

What to do about it

1. Check whether any teen-age users in your household or organization have a ChatGPT account, and confirm whether the teen experience has been applied automatically.
2. Do not rely solely on platform defaults. Link accounts through OpenAI’s parental controls as a second layer, even if automatic protections are active.
3. If you are deploying AI tools for any user-facing product, document what safety evaluations the model vendor has published and note what methods are missing before you ship.
4. Watch for independent audits of ChatGPT for Teens from Common Sense Media or Stanford Medicine, which were involved in pre-launch testing and may publish follow-up findings.

The takeaway: automatic defaults are a genuine improvement, but they are only as good as the detection rate OpenAI has so far declined to publish.

Source: Bing News · OpenAI