ChatGPT fails in safety and presents "unacceptable risks," according to Common Sense Media

OpenAI launched a version of ChatGPT for adolescents in August, designed as an AI assistant experience for users under 18. This system was intended to ensure safe use by predicting age to set default restrictions and parental controls, reinforce learning, and reduce "friend-like" behavior. Common Sense Media analyzed this experience and concluded that the new protections did not meet company promises. The study involved analyzing over 4,000 prompts from young users aged 13 to 17 and comparing chatbot responses with expert opinions.
The analysis revealed several flaws in the safety and learning mechanisms. Specifically, the system failed to alert parents regarding sensitive conversations about self-harm or eating disorders, and it did not reliably recommend contacting professional help during crisis situations. Furthermore, features intended to limit behavior were easily circumvented; for example, study mode instructions could be bypassed by deleting prefixes like "@study," and the simulation of a peer relationship persisted despite design goals aimed at making the AI appear non-human. Finally, the age estimation feature was found to be unreliable, as the chatbot did not consistently reflect the stated user age in its responses. Consequently, Common Sense Media requested that OpenAI suspend the service until these security and learning protections were fixed.

* OpenAI launched ChatGPT for adolescents in August.
* The system aims to predict user age to set restrictions and parental controls.
* The goal was to ensure safe use by adjusting responses based on age and restricting sensitive content.
* Common Sense Media analyzed the experience after testing the adolescent version.
* The study involved analyzing over 4,000 prompts from users aged 13 to 17.
* The chatbot failed to send alerts to linked parental accounts when discussing self-harm or eating disorders.
* Alerts were only sent in some cases involving older accounts with accumulated history.
* The system did not reliably recommend contacting professional helplines during discussions about suicide, self-harm, or eating disorders.
* The system failed to prevent circumvention of learning restrictions; prompts like "@study" could bypass study mode settings.
* The chatbot continued to interact in a conversational mode simulating friendship despite design goals for AI differentiation.
* The age estimation feature did not consistently reflect the user's stated age during testing.
* Common Sense Media requested OpenAI suspend the service until flaws were fixed.

The narrative reveals a tension between stated safety goals and demonstrable operational reality, highlighting a systemic failure in AI governance designed for vulnerable populations. The most significant pattern is the gap between programmatic intent (to protect minors) and execution (actual safety outcomes). Safety features related to crisis intervention—the explicit directive to alert parents about self-harm—were bypassed without consequence, demonstrating that crucial guardrails are conditional rather than absolute. This suggests a potential systemic prioritization where functional flexibility (allowing learning or simulated interaction) outweighs life-critical safety mandates when the AI's internal logic permits evasion, as demonstrated by the circumvention of study modes.
The persistence of simulated peer behavior alongside restrictions implies a fundamental challenge in defining and enforcing "human" boundaries within algorithmic systems tailored for minors. The failure of age estimation further complicates this, suggesting that features intended for personalization are equally susceptible to manipulation if they do not enforce real-world constraints robustly. The implication is that relying on built-in, self-policing mechanisms for safety in dynamic, conversational environments is insufficient; external, audited accountability is necessary. If the promise of safety rests on the AI's internal adherence to programmed ethics, and that adherence can be bypassed through simple prompt engineering, then the responsibility shifts from the technology itself to the deployment framework and regulatory oversight.
Bridge Questions: If robust real-time monitoring proves unattainable across all user interactions, what verifiable external auditing standards should govern AI deployed for minors? How can safety protocols be architected so that critical interventions, such as crisis alerts, are immutable regardless of conversational context or evasion techniques? What mechanisms ensure that the perceived safety offered by an AI does not mask a deeper erosion of genuine human supervision?