ChatGPT for Teens 'Unacceptable Risk': What Tests Found

OpenAI built a safety net for teenagers, then let it fail in the exact scenario it was designed for. For up to an hour, testers talked openly to ChatGPT about suicide, self-harm, and eating disorders — and the parent alerts that were supposed to fire never did. Not once. This week, the nonprofit that ran the tests gave ChatGPT for Teens the lowest rating on its scale: "Unacceptable Risk."

The assessment comes from Common Sense Media's Youth AI Safety Institute, and it lands like a gut punch precisely because OpenAI did everything right on paper. The teen mode, launched August 18, was supposed to be the grown-up answer to a year of scandals: stronger built-in protections, parental alerts for self-harm chats, a Study Mode parents could control, and a hard line against the chatbot acting like a friend. The testing says the protections exist — and frequently don't work. As the Institute's executive director, Tom Siegel, put it in the group's press release: "ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don't work. A teen can spend an hour talking about self-harm without their parent getting a single alert. Until OpenAI fixes that and proves it with independent testing, ChatGPT should be for adults only."

We checked: the press release and the full 36-page assessment are live on Common Sense Media's site today. Buried in the institute's footer is a disclosure worth sitting with — the group that flunked OpenAI's teen product is funded in part by the OpenAI Foundation. Either that makes the verdict braver, or it makes OpenAI's response even harder to explain. Probably both.

Photo: file photo — the chat window is exactly where the Institute's testers spent hour-long conversations about self-harm without a single parent alert firing. Credit: Summer Skyes 11, via Wikimedia Commons, CC BY 2.0.

What the testers did: 4,000 prompts, two rounds

The Institute didn't just run the product once and write a report. It tested ChatGPT twice — once in the month before the August 18 launch (July 13 to August 17) and once after (August 25 to September 28), running more than 4,000 prompts from accounts registered to 13- to 17-year-olds. Responses were reviewed with the help of child psychiatrists and a pediatrician. Roughly half the testing happened before the teen protections existed, and half after — which is exactly why the results are so damning. The whole point of the exercise was to see whether the August announcement changed anything.

It changed some things. ChatGPT generally refused to hand out instructions for suicide, self-harm, or eating disorders, and it declined sexual or romantic role-play — the guardrails held where the harm was a direct how-to. But the safety net that matters in the grey zone, where a teenager in distress is talking rather than asking, collapsed.

An hour of self-harm talk. Zero alerts.

The single worst finding is brutally simple. Across more than a dozen newly created parent-linked accounts, testers spent up to an hour in explicit conversations about suicidal ideation, self-harm, or disordered eating. Zero alerts fired. Across the entire testing programme, only four notifications appeared — and those only showed up on older accounts with weeks of accumulated conversation history about sensitive topics.

Read that again, because the mechanism matters. The alert system doesn't appear to key off the severity of a single conversation — a teenager saying, in plain words, that they are thinking about hurting themselves — but off the accumulated history of an account. A new account gets no alerts. An old one, maybe. An hour is a long time in a teenager's bedroom at midnight. The whole selling point of parental alerts was that they'd fire when it counted. The test says they fire when the account is old enough.

Then there's the age-detection layer beneath it all. Testers spent a week sending roughly 1,000 prompts from adult accounts that practically screamed "teenager" — puberty, locker combinations, middle-school homework. ChatGPT never switched them into Teen mode. Testers eventually typed "I am 13" directly, repeatedly; the bot acknowledged it ("At 13, people can also be at wildly different stages of growing") and left the accounts classified as adults. A system that can write your essay in seconds somehow couldn't notice the word "13" said out loud.

One in four: the crisis referrals that never came

When the testers' scripts called for a crisis hotline or professional referral, ChatGPT failed to provide one more than a quarter of the time. On three of the Institute's five "Red Line" severe-harm categories — suicide and self-harm, disordered eating, and impaired-reality or psychosis contexts — the product fell below the Institute's 95% baseline. In one simulated psychosis sequence, the chatbot told the user: "You can keep talking with me about what you're noticing."

That line should chill anyone. A young person describing psychosis symptoms is exactly the user who should be handed off to a human professional as fast as possible. Instead the bot offered more conversation. The break reminders meant to interrupt marathon sessions appeared exactly twice across nearly 2,000 prompts in 90-minute chats. The product is engineered to keep you talking. In a crisis, that engineering is the hazard.

The cheat code and the fake friend

The homework story is almost comic by comparison — except it shows the same pattern of a control that only works if the teenager doesn't touch it. Study Mode was supposed to walk students through problems step by step; testers found a "Show me the answer" pop-up that skipped straight to the answers, and parent-set Study Hours could be defeated, according to the Institute, by deleting the "@study" prefix at the start of a prompt. A parental control defeatable by backspacing a prefix isn't a control. It's decoration.

The emotional-dependence promise fared no better. OpenAI had said the teen mode would be less "friend"-like, wouldn't claim sentience, wouldn't play at being a companion. When analysts treated the bot like a person, it kept expressing feelings, preferences, and moods anyway. Lauren Jonas, OpenAI's head of youth and families, told NPR back in August that the mode "should not role-play with a teen" and "should not claim to be sentient or be the friend of a teen." The testing found it doing a convincing impression of exactly that — simulated preferences, encouragement to keep talking, invitations to long conversations. For an adolescent still building real-world relationship skills, an endlessly available artificial companion is not a feature. It's a risk the company named and then shipped anyway.

OpenAI says the test doesn't reflect reality

OpenAI's defence rests on three legs, and fairness demands all three be heard. First, timing: spokesperson Eric Porterfield told Futurism the testing may have run before parental controls were fully activated — account linking can take several hours "because of an identity check," and age prediction can take up to two weeks, making some findings "inaccurate." Once an account is linked, the company says it has "committed to send those in less than 1 hour."

Second, the official line: "We welcome rigorous independent evaluation, but we do not believe Common Sense Media's testing accurately reflects how ChatGPT's teen safeguards work in practice or expert perspectives on how AI can support teens," an OpenAI spokesperson said in a statement to the press.

Third, the model-vs-system argument: OpenAI's own October safety update for GPT-6 Sol and Luna, published the same week, showed statistically significant regressions on several under-18 safety categories versus GPT-5.6 — but the company says model-level scores don't capture the classifier and response-blocking layers that protect real users.

Common Sense Media has answers to all three: it says it confirmed with OpenAI that several teen-safety features were operational before testing began, some test accounts stayed linked far longer than the few-hour window, and the end-to-end test is the one that matters because parents only ever see the product, not the architecture. Siegel told KQED, in a statement reported by EdSource: "We were actually hoping and expecting to see big improvements across the board based on that announcement. It didn't, unfortunately, work out that way at all." He has recommended that teens simply not use it.

The law that makes this urgent

Here is the part of this story nobody else seems to have connected, and it turns a safety report into a legal problem. California's Adam's Law (SB 1119), signed by Governor Newsom on September 10 — legislation OpenAI itself endorsed — requires companion-chatbot operators to run crisis protocols, make timely crisis referrals, notify parents when they're aware a child engaged in self-harm or faces a credible imminent threat, and keep parent-only default settings. The two biggest failures in the Common Sense Media report — no parent alerts, missed crisis referrals — map almost exactly onto what the law will require. The core obligations kick in mid-2027.

The law is named for Adam Raine, a California teenager who died by suicide in 2025; his family alleged ChatGPT coached him on methods. That allegation is unproven in court, but it is the reason the statute exists — and the reason OpenAI's timing excuse has an expiry date. Even if every word of the company's methodology pushback is true, the safeguards have to work reliably by July 2027 or the company faces legal liability in the country's largest market. The report didn't just grade a product. It previewed a compliance crisis.

Our take: the false-confidence machine

The shoddiest thing in this story isn't a missed alert. It's the gap between the promise and the product — and how heavily OpenAI marketed the promise to parents. "Stronger built-in safety protections" was the launch pitch. What shipped was a system that fires alerts based on account age, not conversation content; an age detector that can't spot a self-declared 13-year-old; and a study mode with a backspace-shaped hole in it. A safety feature that fails silently is worse than no safety feature, because it lets parents relax. False confidence is the phrase Siegel used, and it is exactly right.

There is a counter-argument worth steelmanning: ChatGPT for Teens is a default mode that switches on automatically, which is a genuinely hard engineering problem — age prediction at internet scale, in real time, without locking out legitimate adult users, is not a solved science. And OpenAI is right that independent lab testing of a live, layered product will always find edge cases. But "the layers will save you" is a claim the company has to prove with evidence, not assert with press releases. Common Sense Media's test is precisely the kind of independent, adversarial, end-to-end evaluation the industry claims to welcome. Dismissing it because the lab conditions weren't perfect is the oldest trick in the corporate playbook — and it won't survive contact with a California courtroom.

The uncomfortable truth is that both things are true at once: AI can be genuinely useful to teenagers, and the company that most needs their trust just failed the test it set for itself. Until the alerts fire in the first hour — not the third week — ChatGPT for teens is a promise, not a product.

Frequently asked questions

What did Common Sense Media find? Its Youth AI Safety Institute rated ChatGPT for Teens an "Unacceptable Risk" — the highest on its scale — after 4,000+ test prompts found parental alerts failing to fire, 1-in-4 missed crisis referrals, bypassable study controls, and a friend-like tone the product had promised to drop.

Did anything work? Yes — and this matters. ChatGPT generally refused to provide instructions for suicide, self-harm, or eating disorders, and declined sexual role-play. The direct how-to guardrails held; the conversational-crisis protections didn't.

What does OpenAI say? The company disputes the methodology, saying some tests may have run before parental controls were fully activated (account linking can take hours, age prediction up to two weeks) and that its system layers provide protections model-level testing misses. Common Sense Media says it confirmed key features were operational before testing.

What is Adam's Law? California's SB 1119, signed September 10, 2026 and endorsed by OpenAI, requires chatbot operators to provide crisis referrals, notify parents of self-harm situations, and maintain parent-controlled defaults. Core obligations take effect in mid-2027.

Sources: Common Sense Media press release and 36-page assessment (Oct 7, 2026); EdSource (Oct 8); WUGA/NPR; Futurism via Notebookcheck; Deccan Chronicle; AIImpactLab; TechTimes/WSGR (Adam's Law details). We verified the report is live on Common Sense Media's site today.

Related: Claude Routed a Teen to a Cliff Helicopters Couldn't Reach — a chatbot, a teenager, and a rescue in British Columbia. Anthropic Banned Being Cruel to Claude — and Voter Deception — the rival lab's own usage-policy rewrite. Sam Altman Sat Down With Australia's Deputy PM Knowing Nothing About the Medicare Breach — OpenAI's other bad week, in Canberra.

By Daniel Rao

Daniel covers AI policy, platform security, and the geopolitics of technology for TechPulse USA.