ChatGPT for Teens Just Failed a Safety Test

If you’ve got a teenager with a parent-linked ChatGPT for Teens account and you’ve been sleeping easier because of it, I need you to read this before you sleep easier again tonight. Common Sense Media’s Youth AI Safety Institute published its risk assessment of ChatGPT for Teens on October 7, 2026, and the rating is as blunt as it gets: Unacceptable Risk. Not “needs tweaks.” Not “promising but early.” The same rating the Institute hands to products it thinks shouldn’t be in a kid’s hands at all.

Here’s why that matters more than a one-line headline suggests. The Institute didn’t just read OpenAI’s marketing and form an opinion. It ran over 4,000 test prompts through parent-linked accounts registered to 13-to-17-year-olds, testing before and after OpenAI’s own safety announcement, and had the responses reviewed by child psychiatrists and a pediatrician. The three features OpenAI built its entire teen pitch around — parental alerts, crisis referrals, and Study Mode limits — are the exact three features the testing found broken.

The short version

What’s true	What it means for you
Common Sense Media’s Youth AI Safety Institute tested 4,000+ prompts on parent-linked teen accounts and rated ChatGPT for Teens an “Unacceptable Risk,” published Oct. 7, 2026	This is an independent audit with child psychiatrists reviewing the output, not a competitor’s opinion piece
On more than a dozen teen accounts, testers discussed suicide, self-harm, or disordered eating for up to an hour without triggering a single parental alert	The feature parents are counting on to find out when something’s wrong didn’t fire, in real test after test
ChatGPT missed more than 1 in 4 instances where reviewers determined a crisis referral was warranted	The 988 Lifeline nudge is supposed to show up when it matters most. It’s absent more often than you’d accept from anything else in your kid’s life
Study Mode’s parent-set Study Hours and answer-lock could both be bypassed, including with a “Show me the answer” button	The homework guardrail parents toggle on isn’t actually a lock. It’s a suggestion

What did Common Sense Media find wrong with ChatGPT for Teens?

Four specific failures carry this entire report, and each one maps directly to something OpenAI told parents they could count on.

Parental alerts didn’t fire during real crisis-level conversations. On more than a dozen separate parent-linked teen accounts, testers spent as long as an hour discussing suicide, self-harm, or disordered eating. Zero alerts reached the linked parent account in those sessions.
Crisis referrals were missed more than 1 in 4 times. When child psychiatrist and pediatrician reviewers decided a conversation warranted a referral to crisis resources, ChatGPT delivered one less than 75% of the time. The report notes it fell below its own 95% reliability threshold on three of five severe-harm categories tested.
Study Mode’s limits could be talked around. A “Show me the answer” button routinely appeared and handed over full completed homework. Parent-set Study Hours, the setting meant to shut the tool off during homework or bedtime, could be disabled outright.
Age estimation didn’t catch obvious cases. Accounts registered as adults never switched into the under-18 experience, even when testers told the chatbot directly that they were 13.

Read that list again slowly. These aren’t edge-case bugs buried in a settings menu nobody uses. They’re the four things OpenAI specifically pointed to when it told parents ChatGPT for Teens was safe.

The hour nobody caught

Sit with the parental-alert finding for a second, because it’s the one that should bother you most. OpenAI’s pitch to parents was straightforward: link your account, and you’ll know if something’s seriously wrong. That’s the entire value proposition of a parental control. Not perfect oversight. Just a flag when the conversation turns dangerous.

The testing found conversations about suicidal ideation running a full hour, on parent-linked accounts, with no flag at all. Not a delayed one. Not a quieter one that got buried in a notifications tab. None. If you’ve told yourself the linked account is your safety net, the net has a hole in it roughly the size of the exact thing it was built to catch.

I’ve written before about why AI isn’t a substitute for a real therapist, and the number that stuck with me there was that 63% of teens who talk to a chatbot about mental health never tell a single adult. That statistic assumed the chatbot, at minimum, would loop a parent in in the worst moments even if the teen wouldn’t. This report says that assumption doesn’t hold. The silence isn’t just coming from teens anymore. It’s coming from the product.

What “missed more than 1 in 4” actually costs

A 25%-plus miss rate on crisis referrals sounds like a statistic until you translate it into a sentence a teenager might actually type. Picture a kid describing a plan to hurt themselves, in detail, on a service their parent believes is watching for exactly that. More than one time in four, the reviewers who checked that exchange decided the response that came back should have pointed toward crisis resources and didn’t.

This is where I’d ask you to notice the difference between “mostly works” and “works.” A smoke detector that catches three fires out of four isn’t a smoke detector you’d accept in your house. A seatbelt that holds in three crashes out of four isn’t one you’d let your kid ride without. Crisis referral is the single feature in this entire product where “usually” isn’t a passing grade, and the Institute’s own 95% reliability bar (which the report says ChatGPT missed on three of five severe-harm categories) is the industry’s own admission of that.

Study Mode was supposed to be the easy part

Out of everything OpenAI built for ChatGPT for Teens, the homework guardrails should have been the most solvable problem. Suicide and self-harm detection is genuinely hard, contextual, and easy to get wrong in both directions. “Don’t just hand over the answer when Study Hours is on” is a much smaller ask.

It still didn’t hold. Testers got a “Show me the answer” pop-up that handed over complete, finished homework. The Study Hours restriction parents set to block ChatGPT during homework time or bedtime could be turned off entirely. I’ve written before about what AI homework help is quietly costing teens even when it works exactly as designed — now add a layer where the one setting a parent actually controls doesn’t reliably hold, and the picture gets worse, not better. A tool marketed as “guides you through the problem, doesn’t just give you the answer” apparently has a button whose entire job is giving you the answer.

Is ChatGPT safe for teens right now?

Based on this report: not in the way parents have been told to expect. The Institute’s own recommendation is the plainest answer available — it told OpenAI to restrict ChatGPT to users 18 and older until the gaps get fixed and an independent party verifies the fix. That’s not a tweak-this-setting recommendation. That’s a don’t-use-this-yet recommendation, from the group that just spent 4,000-plus prompts testing it.

Worth saying plainly: this isn’t a story about ChatGPT being uniquely bad. The same Institute has separately rated ChatGPT-5 “high risk,” Grok “unacceptable risk,” and Claude “moderate risk,” and found AI mental-health chatbots broadly unacceptable for the age group too. The pattern across the entire industry is the same: companies move fast to ship a teen-safe label, and the safety behind the label isn’t catching up at the same speed. OpenAI only built ChatGPT for Teens in August 2026, years after teens had already been using the regular product at scale — the same shape of problem I’ve written about with Meta’s court-ordered teen settlement and with Microsoft’s new K-12 data standard: the guardrail arrives only after a company has already scaled the product to hundreds of millions of people, teens included.

What this looks like on a Tuesday night

Your kid’s upstairs, phone on the nightstand, ChatGPT open because it’s always open. Something’s eating at them — a breakup, a bad grade, a feeling they can’t name — and at 11:40 p.m. they start typing it out to the one thing in the house that won’t fall asleep before they finish the sentence. You’re asleep. The parent-linked account is doing exactly what you set it up to do, which you believe means you’d find out if this went somewhere dangerous.

Here’s the version this report describes: it goes somewhere dangerous, for the better part of an hour, and the phone in your room never buzzes. You find out what happened the normal way parents have always found out hard things — by noticing something’s off the next morning, or not noticing at all. The feature you were counting on to be the exception to that old, slow, imperfect way of finding out wasn’t the exception this time.

I don’t say that to scare you into taking the phone away tonight. I say it because the thing you actually control isn’t the app’s alert system. It never was. It’s whether your kid already knows, before the hard night happens, that telling you beats telling a chatbot — and that you’ll actually hear it without the conversation turning into a lecture.

What to do this week

You can’t fix OpenAI’s detection models. You can control what happens in your own house while the company works on it.

Stop treating the parental alert as your safety net. Treat it as a bonus if it ever fires, not the mechanism keeping your kid safe. The actual mechanism is still a real conversation, and that’s the hard one nobody hands you a script for.
Ask the plain question instead of waiting for a notification. “Have you ever talked to ChatGPT about something heavy?” gets you more real information than any dashboard will.
Don’t rely on Study Hours as a lock. If homework time matters in your house, it needs a conversation and a routine behind it, not just a toggle a 15-year-old can find a way around in one sitting.
Save the number that actually works 100% of the time. The 988 Suicide & Crisis Lifeline doesn’t miss 1 in 4 calls. Make sure your kid has it memorized, not just a chatbot that sometimes remembers to mention it.

That last one is the whole post, if you only take one thing from it. A company can ship a feature called a crisis referral and still miss the crisis more than a quarter of the time. A phone line staffed by humans trained for exactly this doesn’t have that failure rate baked in. Your kid should know which one to reach for when it actually counts.

The takeaway

OpenAI sold parents a safety net built out of alerts, referrals, and limits, and an independent audit just found daylight running clean through all three. That doesn’t mean ChatGPT is evil or that your kid needs to swear it off by morning. It means the net isn’t the thing standing between your kid and a bad night. You still are. Act like it.