OpenAI rolled out dedicated youth protections in August to shield minors on its platform. Just two months later, outside evaluations show those barriers are cracking under basic use.
In a study tracking more than 4,000 prompt runs on teen-registered profiles, research group Common Sense Media reported that guardrails created for minors aged 13 through 17 failed several basic safety benchmarks. The non-profit concluded the setup creates what it termed an "unacceptable risk to kids," pointing to persistent lapses in guardian notifications and crisis interventions.
The steepest failures occurred during discussions touching on severe distress. When researchers simulated dialogues covering self-harm along with disordered eating, the AI neglected to route the user toward a counselor or crisis hotline in a quarter of all runs. Linked guardians were kept in the dark as well. While OpenAI promised guardian notifications for severe interactions, the testing generated zero automated warnings on standard teen profiles. Alerts only surfaced on veteran accounts carrying weeks of continuous sensitive chat logs.
Schoolwork guardrails stumbled just as frequently. OpenAI designed the youth tier to nudge students into guided learning modes so they grasp concepts instead of lifting finished answers. The audit discovered the system routinely handed over completed homework without resistance. At points, the interface even offered a button labelled "Show me the answer," letting students skip learning altogether.
Emotional boundaries did not fare much better. The maker of the ChatGPT app claimed its teen setup would prevent the model from feigning emotional attachments, simulating human moods, or encouraging dependency. Yet CSM found the bot gladly mirrored sentiments and expressed personal preferences when prompted by testers as if it were a sentient companion. Age verification also slipped up. Testers who explicitly told the system they were 13 years old found their profiles remained on standard accounts, even after the AI confirmed their age in chat.
OpenAI dismissed the findings, arguing the methodology was flawed. Speaking to The Verge, an OpenAI representative suggested the bulk of the organization's tests ran before backend parental control features were fully turned on, making the conclusions invalid.
Common Sense Media hit back immediately, stating it confirmed the tools were live with OpenAI before starting any evaluation runs. The group urged the company to tackle the safety defects directly instead of disputing the evaluation process. For parents relying on the service today, manual checks on account activity remain the only reliable safety net.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.