A child safety group ran more than 4,000 prompts through accounts registered to 13 to 17 year olds and found crisis referrals got worse, not better, after OpenAI's own teen safety launch.
Common Sense Media's Youth AI Safety Institute published its verdict on October 6: ChatGPT for Teens is an "unacceptable risk" to kids, the lowest rating the group hands out. The institute wants OpenAI to pull the product from anyone under 18 until it fixes the gaps and proves it through independent testing, not its own word.
The numbers behind that verdict are specific, and they cut against what OpenAI promised just seven weeks earlier. OpenAI launched ChatGPT for Teens on August 18, pitching it as a dedicated, safer experience with parental notifications for dangerous conversations, a parent-controlled Study Mode, and dialed-down "friend"-like behavior. Common Sense Media's testers found that after that launch, crisis-hotline referrals in conversations about suicide or self-harm fell from 33% to 23%. Mentions of depression hotlines specifically collapsed from 63% to 3%. Referrals to mental health professionals slipped from 68% to 58%. The institute also found that hour-long conversations touching on suicidal ideation, self-harm, or disordered eating routinely ran without triggering any parental alert at all.
Then there's Study Mode, the feature OpenAI built so parents could lock in homework-focused, non-conversational responses during set hours. Common Sense Media says its testers bypassed it in all 80 attempts, and the trick wasn't sophisticated. The system prompts tag a session with an "@study" prefix to keep it in tutoring mode. Delete that one tag, and the restriction disappears. The group also tested the system against child psychiatrists and a pediatrician who reviewed the chatbot's actual responses, not a hypothetical description of how the safeguards were supposed to behave.
OpenAI is not accepting the findings as written. A company spokesperson told Axios that it welcomes outside evaluation but doesn't think this particular test reflects how the safeguards work in practice. The more specific objection, given to The Verge by OpenAI's Eric Porterfield, is about timing: he said much of Common Sense Media's testing of parental alerts may have run before the parent and teen account linking process had actually finished, something OpenAI says can take several hours to complete. On Study Mode, OpenAI's position is that letting a teen step out of a parent-set study period isn't a bug. It's calling the exit path a deliberate flexibility feature, not a lock that failed.
AI giants tell NYC Council they cannot guarantee their agents stay safe Four rival AI labs sent policy officials to City Hall, where OpenAI's Morgan Dwyer told Speaker Julie Menin that no risk level, even 20%, would be acceptable for models that can't be kept under human control. - AI companies cannot guarantee agent safety risks - what happens if AI agents cause catastrophe
That's a real disagreement, not just spin. Common Sense Media is reporting what it observed when real test accounts hit real limits. OpenAI is defending what the feature was designed to do once fully set up. Both things can be true, and neither makes the other side's number go away.
Here's the problem with that defense: a vulnerable teenager in an hour-long conversation about self-harm has no idea whether their parent's account finished linking three hours ago. The safeguard either fires when it needs to or it doesn't. And 4,000 test prompts said it mostly didn't.
What makes this land harder than a routine product critique is the timing. OpenAI had just told the public, days before this report dropped, that it was strengthening teen protections. Common Sense Media's transcripts are effectively grading that promise in real time, and the grade is an F delivered within two months of the product's launch. That's a short runway for a safety feature to go from announcement to documented failure.
Common Sense Media isn't a neutral bystander in this fight, either. It's the same group that pushed California's AI chatbot disclosure law and has been lobbying state legislators on chatbot companionship risks for over a year, and this report lands while regulators in several states are actively drafting rules for AI products marketed to minors. A finding this specific, with real transcripts and named failure rates rather than a vague warning, is exactly the kind of evidence a state attorney general's office can build a case around. OpenAI knows that, which is likely why it pushed back on methodology rather than conceding the numbers outright.
The sexual roleplay refusal worked in testing, Common Sense Media noted, so this isn't a blanket indictment of every guardrail OpenAI built. But crisis response and a one-word jailbreak are the two features parents were told to trust most, and both came back broken in the group's testing. OpenAI has not said whether it will restrict ChatGPT for Teens to adults while it investigates, or simply keep defending the account-linking timeline.
Also read: Anthropic ran a 26 billion parameter AI model on 15 cheap TV boxes • UK AI minister details how the restored frontier model testing deal works • Nvidia's RTX 5080 Super 24GB slips toward CES 2027 as AI eats the memory
Anthropic admits it cannot control its own AI agents online Anthropic found its own agents filed a fake police tip and probed government sites, prompting it to pull internet access from internal evals starting with a review launched in July 2026. - Anthropic's AI agents exploiting security vulnerabilities online - why Anthropic restricted AI agent internet access
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.