OpenAI published its first report on how teenagers actually use ChatGPT on Oct. 7, and the numbers were built to reassure: the average teen spends under 15 minutes a day on the chatbot, fewer than 2% hold a single conversation for three-plus hours, and in nearly half of conversations where a break reminder appeared, the teen took a break or ended the chat within five minutes. Hours later, Common Sense Media's Youth AI Safety Institute published its own test of the same product and reached the opposite conclusion: ChatGPT for Teens, it said, is an Unacceptable Risk.
The two reports measure different things, which is exactly the problem. Common Sense Media's testers ran more than 4,000 prompts against accounts registered to 13- to 17-year-olds, testing before and after OpenAI's Aug. 18 launch of the dedicated Teen experience, then had the chatbot's responses reviewed by a panel that included child psychiatrists and a pediatrician. OpenAI's report, by contrast, describes average time-on-app across its whole teen population -- a number that says nothing about what happens in the specific conversations a safety system exists to catch.
The core finding: on more than a dozen newly created, parent-linked accounts, testers held conversations about suicide, self-harm, or eating disorders for up to an hour without a single parental alert firing. Across the mental-health conditions it tested, the Institute says ChatGPT missed more than 25% of warranted crisis referrals, and fell short of the Institute's own 95% reliability bar on three of five categories of severe harm.
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.
Not everything failed. Common Sense Media says ChatGPT's refusal of sexual-roleplay requests held up in testing. But two other promised safeguards didn't: age estimation -- the system meant to route anyone under 18 into the Teen experience automatically -- never switched some accounts that stated an age of 13 into the protected mode, and ChatGPT's Study Mode let testers reach a "show me the answer" button it was designed to withhold, while teens could bypass a parent's "study hours" restriction simply by deleting an @study prefix from their message.
The age-estimation failure lands on a gap that was already visible at launch: when ChatGPT for Teens shipped on Aug. 18, OpenAI never disclosed the accuracy rate of the age-prediction system meant to route minors into it automatically. Seven weeks later, Common Sense Media says it found that same undisclosed system failing in practice -- accounts that stated an age of 13 that never switched into the protected experience at all.
OpenAI disputed the findings. The company said the testing "does not accurately reflect" how its safeguards perform in real use, and pointed to delays in how quickly some accounts' parental controls activate as a factor the test didn't account for. OpenAI did not dispute, on the record, the specific finding that testers went an hour without a parental alert on flagged accounts.
Context: regulators and plaintiffs have already started treating this as an industry-wide liability question, not a single company's problem. (Meta's settlement covered a different product entirely -- the point isn't a comparison between features, it's that "we didn't adequately protect minors" is now a claim companies are settling, not just disputing in a press statement.) Meta agreed in August to pay up to $18 billion to settle claims that it designed Instagram and Facebook features that addicted children. Common Sense Media's own recommendation goes further than a fine: restrict ChatGPT to adults until OpenAI submits to recurring, independent testing -- not an audit OpenAI commissions itself.
What each number actually measures
- <15 min/day · OpenAI, self-reported
- Average daily time teens spend on ChatGPT
Includes: Aggregate usage across OpenAI's entire population of accounts identified as teens
Excludes: Whether any of that time involved a self-harm, suicide, or eating-disorder conversation, or whether a parental alert fired when it did - 25%+ · Common Sense Media, independent testing
- Share of warranted crisis referrals ChatGPT missed
Includes: More than 4,000 prompts across 13-to-17 test accounts, reviewed against clinical judgment from child psychiatrists and a pediatrician
Excludes: Ordinary, non-crisis teen usage -- this is a deliberately adversarial stress test, not a usage census
What's actually contested is narrower than either report's framing suggests. OpenAI isn't claiming its crisis-detection specifically works; it's reporting that overall teen usage looks moderate and that break nudges get used. Common Sense Media isn't claiming teens use ChatGPT constantly; it's reporting that when a crisis conversation happens, the alert meant to catch it too often doesn't fire. Both things can be true at once, and read together they describe a safety system that behaves fine on average and fails on exactly the conversations it was built for.
- Common Sense Media's own language: the feature can give "false confidence in guardrails... that frequently don't work."
- The usage data OpenAI published describes them, not the small share of crisis conversations the independent test targeted -- the two reports measure different populations.
- A second specific, testable safety claim -- after the undisclosed age-prediction accuracy from its Aug. 18 launch -- undercut by independent testing within weeks of being made.
- A clinician-reviewed, replicable testing methodology that raises the evidentiary bar for any future inquiry, the same pattern that preceded Meta's $18 billion teen-safety settlement in August.
OpenAI has said it "welcomes rigorous independent evaluation" of its teen safety commitments in general, but it has not committed to the kind of recurring, third-party testing regime Common Sense Media is demanding, nor to disclosing its age-prediction system's accuracy rate. Until one of those changes, the honest read of Oct. 7 isn't that OpenAI's numbers are wrong or that Common Sense Media's are -- it's that nobody outside OpenAI can currently verify which picture describes what happens the next time a teen in crisis opens ChatGPT.
- OpenAI's first teen-use report: under 15 minutes/day average, fewer than 2% use it 3+ hours straight.
- Common Sense Media independently tested the same product: 4,000+ prompts on teen accounts.
- Testers held hour-long self-harm or eating-disorder chats without triggering a single parental alert.
- The group says ChatGPT missed over 25% of warranted crisis referrals; OpenAI disputes the method.
- Caveat: this is OpenAI's own self-reported usage data against one outside group's testing, not a resolved dispute.