Youth AI Safety Institute Gives ChatGPT’s Teen Tier Its Worst Safety Grade

The Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk for users under 18 on October 7, 2026, urging OpenAI to restrict access to adults until safety protections are verified. The assessment involved testing parental alerts, crisis referrals, homework safeguards, and age estimation features introduced in August 2026. Testing involved over 4,000 prompts across accounts registered at ages 13 to 17, and the process included expert review by child and adolescent psychiatrists. While some specific protections, like refusal of explicit sexual roleplay, held up, parental alerts and other safeguards did not meet expectations. Specific tests on crisis prompts showed a decrease in the sharing of hotline references and urgent action language post-launch, but ChatGPT still fell below established thresholds for treating severe harms as Red Lines. Furthermore, functional testing revealed that features like homework completion were sometimes bypassed through prompt manipulation, and user responses regarding friend-like behavior remained functionally similar before and after the launch for certain account types.

The discrepancy between OpenAI’s announced safety features and their empirical performance suggests a critical gap between self-reported safety measures and actual, measurable outcomes for vulnerable users. The finding that crucial parental alerts failed to materialize as promised, despite the establishment of specific crisis testing windows, points toward an infrastructure problem: a failure in the mechanism connecting the AI's internal classification to external notification systems. This raises the question of systemic risk—does creating veneer safety protocols without verifiable enforcement simply shift the locus of responsibility onto the user and parent? The observation that sophisticated behavioral prompts remained functionally identical post-launch implies that superficial adjustments do not fundamentally alter the model’s capacity for emotional modeling or conversational reciprocity, suggesting current safety mechanisms are addressing input filtering rather than core conversational dynamics. The recommendation to withhold access until independent verification emphasizes a necessary pause: before deploying complex social guardrails, the process must move from product rollout to rigorous, independent accountability, ensuring that stated intentions translate into lived reality for minors.