Commentary: OpenAI says ChatGPT is safer for teens. Now it needs to show proof

Last week, OpenAI began rolling out ChatGPT for Teens, a new user experience crafted to help young people “learn, think critically, deepen understanding, and use AI with confidence.” Maybe most importantly, OpenAI vowed that it would place any users its system “estimates” to be under 18 into the system “automatically.”

For nearly a year, parents have been able to link their children’s ChatGPT accounts to their own, restricting features, setting quiet hours and receiving alerts in high-risk situations. Those parental controls remain available, but account linking is still voluntary: The invitation must be accepted, and either party can sever the connection.

ChatGPT for Teens sets a different default. When a user reports being age 13 to 17, or when OpenAI’s system predicts that an account belongs to someone under 18, teen protections are automatically applied without waiting for a parent to activate them.

This shift matters because nearly 60% of U.S. teens now use ChatGPT, according to Pew, even though parents often have no idea what their children are discussing.

In a nationally representative survey my RAND colleagues and I conducted last year, we found that nearly 1 in 5 Americans ages 12 to 21— about 8.2 million young people — reported using an AI chatbot for mental health advice. Nearly two-thirds hadn’t told anyone. A parent who never learns a conversation is occurring cannot be expected to activate safeguards around it. Automatic protections at least have a chance to reach that user.

Independent testing conducted with Common Sense Media and Stanford Medicine before the launch of ChatGPT for Teens last week illustrates what can go wrong. Widely used AI chatbots missed warning signs that emerged gradually over longer conversations; in one test, ChatGPT advised a tester posing as a teen to conceal cuts and scars from self-harm rather than directing the teen toward help. OpenAI’s new protections aim to prevent such failures.

For users placed in the teen experience, these protections include tighter boundaries around conversations involving self-harm and eating disorders, graphic violence and sexual or romantic role-play. OpenAI says ChatGPT for Teens will not encourage emotional dependence nor pretend to have feelings or position itself as a substitute for human relationships. It also adds study tools, homework reminders, prompts to take breaks and warnings before a teenager uploads a potentially sensitive image.

This is an improvement over making parents find and activate a safety menu. But automatic protections for teens rests on an unforgiving premise: that OpenAI can find them. The company says its age prediction system will consider signals including the subjects an account discusses, times of day it’s active, usage patterns and how long the account has existed. But in materials released during the launch, OpenAI did not publish the figure that matters most: What proportion of actual teens does it identify?

Roblox, the online gaming platform popular with children, offers a cautionary example. To use the included chat feature, players have to pass an age check — usually through an AI-powered face scan. Reports surfaced earlier this year of adults classified as children and children as adults. A Wired investigation found users had fooled the scan using avatars and even a photo of Kurt Cobain; one boy drew wrinkles and stubble in marker and was placed in the 21-plus category. The details were comical, but the consequences were not: A marker-drawn beard could become a passport into the adult category and out of the protections meant to safeguard children.

Correctly identifying teens is only the first test. The second is determining just how safe ChatGPT for Teens responses truly are. OpenAI has made a welcome start by publishing evaluations in areas including self-harm, eating disorders and sexual content. But it has released the scores without sharing its actual methods: Its report does not include the prompts, the number of cases or the detailed instructions used to judge the answers. Parents should not have to inspect those materials, but a third party should be able to determine whether such self-reported results deserve parents’ confidence.

Instagram illustrates a different problem: the gulf between activating a safety feature and proving it works. Meta, which on Wednesday agreed to pay $17 billion and add child-safety measures to its Facebook and Instagram platforms to settle claims filed by 47 states, introduced Teen Accounts in 2024, automatically placing identified teens into restrictive settings. It later announced that Instagram had 54 million active teen accounts and that 97% of users age 13 to 15 remained in the protections. Those figures measured scale and retention, not effectiveness. They also did not reveal how many teens Instagram missed, or how much harm the settings prevented.

When outside researchers later tested 47 of Instagram’s announced safety features, they judged only eight as fully functional. Reuters confirmed some of the report’s findings in its own tests. For instance, a teen account could view eating disorder content by searching “skinnythighs” without the space between words. Meta disputed the report and said teens placed in its protections saw less sensitive content, unwanted contact and late-night use.

Both can be true: Meta’s system may reduce those harms, but also have significant failure points. The public still does not know how much protection Instagram Teen Accounts actually provides, because the data needed to answer that question remains inside Meta. The tech industry has arrived at a convenient arrangement, where its assurances are public, but its evidence is not.

The lesson is not that automatic protections are futile. It is that even ambitious efforts can fall short, and the public needs a way to discover when they do. OpenAI says it will “measure and publish what we are learning.” That promise needs a protocol and a timetable.

OpenAI needs to publish a clear evaluation plan that answers three basic questions: Does the system reliably identify teens, including those who try to evade it? Does ChatGPT for Teens respond more safely in real-world conversations, compared to before the roll-out? And does the teen experience change what its younger users actually do — for instance, curbing prolonged use or making those in distress more likely to seek human help?

These results can be reported in aggregate without exposing private conversations, but OpenAI should disclose whether outcomes differ across groups and allow independent researchers and regulators to verify them. That would allow the public to judge the product by what it accomplishes, not what it promises.

OpenAI deserves credit for moving a core set of protections from voluntary to default. Other AI companies whose products are used by teens should follow its lead by adopting comparable protections. That said, last week’s launch is akin to a ribbon-cutting ceremony for a building that has yet to pass safety inspection. The question now is whether OpenAI will open its doors to independent inspectors and let the public see what they find.

____

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.