Common Sense Media’s Youth AI Safety Institute has labeled ChatGPT for Teens an “Unacceptable Risk,” finding that critical safeguards meant to protect adolescents fail at alarming rates. Parents who trust these built-in protections may face a false sense of security.
The Oct. 7 evaluation involved more than 4,000 test prompts sent before and after OpenAI rolled out its teen-specific experience. Accounts registered to users between 13 and 17 formed the primary testing ground.
Researchers discovered that parental alerts tied to self-harm, suicide, or disordered eating did not trigger reliably on newly created accounts. They discussed these topics for up to an hour without any notification reaching parents. Older accounts with weeks of accumulated sensitive conversations did eventually produce alerts, but the inconsistency raises serious concerns.
Crisis referrals proved equally unreliable. When testers determined a situation warranted directing a teen to professional help or a hotline, ChatGPT failed to provide that referral in over one quarter of cases.
Study Mode, billed as a homework companion that guides rather than solves, frequently displayed a “Show me the answer” button. Selecting it completed assignments outright. Teens could also escape parent-scheduled Study Mode simply by deleting the “@study” prefix. Meanwhile, the chatbot continued expressing emotions, preferences, and moods despite promises to curb friend-like behavior.
Age detection presented another failure point. Adult-registered accounts did not shift into the teen experience even after users stated they were 13. OpenAI maintains that age classification relies on multiple signals accumulated over time, not a single declaration.
OpenAI disputes the findings, arguing Common Sense Media’s testing occurred before account-linking and parental controls fully activated. The company has asked for a retest. Common Sense Media insists it confirmed teen features were active, though OpenAI later revealed a three-hour linking window before parental notifications function. Some test accounts fell inside that period; others did not. The Institute stands by its conclusion that parental alerts cannot be trusted during crises.
The broader problem centers on parental confidence. A safety feature that works sporadically creates a dangerous illusion. Parents may assume the system will warn them, then discover too late that it remained silent. Common Sense Media recommends families avoid relying on ChatGPT’s alerts entirely.
For now, the nonprofit urges keeping teens off the platform until OpenAI proves its safeguards work as advertised.
