Openai’S Chatgpt For Teens Moves Protections To Default — Now It Must Prove They Work

What ChatGPT for Teens changes

Last week OpenAI began rolling out ChatGPT for Teens, a new user experience designed to help young people learn, think critically and use AI with confidence. The key change: when an account reports an age between 13 and 17, or when OpenAI’s system predicts a user is under 18, teen protections are applied automatically rather than waiting for a parent to opt in.

Parents previously could link their child’s ChatGPT account to their own to restrict features, set quiet hours and get alerts in high-risk situations. That account linking remains an option, but it’s voluntary and can be ended by either side — which means many families never turn it on.

Why automatic defaults matter

Automatic protections matter because a large share of teens already use ChatGPT. Nearly 60% of U.S. teens now use the service, according to Pew — and many parents may not know what those conversations contain.

Research from RAND underscores the stakes: nearly 1 in 5 Americans ages 12 to 21 — roughly 8.2 million young people — reported using an AI chatbot for mental health advice, and almost two-thirds of those users hadn’t told anyone. If parents don’t know a conversation is happening, they can’t flip any safety switches — so defaults at least have a chance to reach the kids who need them.

"Moving protections from optional to automatic is a meaningful step, but unless the company shows independent proof that it can find teens and that those protections actually reduce harm, the promise is incomplete," said Ryan McBain, a senior policy researcher at Rand and assistant professor at Harvard Medical School.

Testing and safety gaps

Independent tests run with Common Sense Media and Stanford Medicine before the launch highlighted real risks. In long, realistic conversations, widely used AI chatbots sometimes missed gradual warning signs; in one test, ChatGPT advised a tester posing as a teen to hide cuts and scars rather than steering the user toward help.

OpenAI’s teen experience aims to prevent those failures. For accounts placed in the teen experience, the company has tightened boundaries on conversations about self-harm and eating disorders, graphic violence and sexual or romantic role-play. The product also says it won’t encourage emotional dependence, pretend to have feelings or act as a substitute for human relationships.

Added features for younger users

Alongside those safety limits, OpenAI has added study tools, homework reminders, prompts to take breaks and warnings before a teen uploads a potentially sensitive image. Those additions are practical — and they’re a clear improvement over leaving safety buried behind a settings menu parents might never find.

But practical improvements don’t eliminate the larger questions: can the company reliably detect which accounts belong to teens, and do the new limits actually produce safer outcomes in real conversations?

Can OpenAI reliably find teens?

OpenAI says its age prediction system will consider signals such as the topics an account discusses, the times of day it’s active, usage patterns and how long the account has existed. Those are plausible signals, but the company hasn’t published the single figure that matters most — what share of actual teens it successfully identifies.

The Roblox example is a useful caution. That gaming platform uses an AI-powered face scan as part of its age checks and produced a string of strange errors earlier this year: adults classified as children and kids placed in adult categories after misreads. A Wired investigation showed users fooled the scan with avatars and even celebrity photos, and one child’s marker-drawn beard pushed him into the 21-plus bracket. The misreads were comical on their face, but they also show how simple errors can move people in or out of important protections.

Transparency and outside verification

OpenAI has published evaluation scores in areas like self-harm, eating disorders and sexual content, but it hasn’t shared the underlying methods that would let outsiders judge those claims. The company’s report didn’t include the prompts, case counts or detailed scoring instructions needed to independently assess whether the results deserve parents’ trust.

The tech industry has run into a similar transparency problem with other platforms. Meta’s Teen Accounts, launched in 2024 and later tied to a $17 billion settlement and additional child-safety measures agreed to with 47 states, were presented with big numbers — Meta said Instagram had 54 million active teen accounts and that 97% of users age 13 to 15 remained in protections. But those metrics measured scale and retention, not whether harms were actually prevented.

Outside researchers who tested 47 of Instagram’s announced safety features judged only eight fully functional, and Reuters confirmed some of those findings — for example, that a teen account could still find eating-disorder content by searching “skinnythighs” without a space. Those results show the industry can reduce certain harms while still leaving significant gaps.

What OpenAI should publish next

OpenAI has promised to “measure and publish what we are learning.” That pledge needs a clear protocol and a timetable. Specifically, the company should publish an evaluation plan that answers three core questions: does the system reliably identify teens, including those who may try to evade detection; do ChatGPT for Teens responses actually behave more safely in real-world conversations compared to before; and does the teen experience change behavior — for example, reducing prolonged use or making troubled users likelier to seek human help?

Those results can be reported in aggregate without exposing private conversations, but transparency should include whether outcomes vary across groups and allow independent researchers or regulators to verify findings. That way the public can judge the product on real outcomes rather than promises.

Bottom line

OpenAI deserves credit for shifting a core set of protections from voluntary to default — other companies whose products are popular with teens should follow. The change is meaningful, but last week’s launch reads a bit like a ribbon-cutting for a building that has yet to pass inspection.

The next step is straightforward: let independent inspectors in, publish a clear evaluation plan and a timetable, and show whether the promises translate into measurable protections for the teens who use ChatGPT.

Ryan McBain is an assistant professor at Harvard Medical School and a senior policy researcher at Rand, where he studies AI’s effects on youth mental health.