ChatGPT for Teens Fails Key Safety Tests, Report Finds

Common Sense Media found ChatGPT for Teens fails in five key areas, including parental alerts and crisis referrals. The assessment raises immediate questions for operators building or moderating AI-assisted youth experiences.

What happened

Common Sense Media has published a formal risk assessment of ChatGPT for Teens and called the mode an “unacceptable risk” for all children under the age of 18. The organization tested the software extensively and found failures in five key areas, publishing the findings in a 36-page report on Wednesday.

The assessment came from the group’s Youth AI Safety Institute, which is funded in part by the OpenAI Foundation. Some advertised protections held up, including the refusal of explicit sexual roleplay, but Common Sense Media said others failed or got worse with the new Teen mode. It warned that the product could give parents false confidence.

“We are concerned that ChatGPT for Teens could give parents false confidence in guardrails that frequently don't work.”

The report was released the same day OpenAI shared usage stats: in one week nearly 1.2 million teens used ChatGPT’s interactive learning visuals to understand math and science concepts, and less than two percent of teen users spent more than three consecutive hours on ChatGPT.

Key facts

• Common Sense Media deemed ChatGPT for Teens an “unacceptable risk” for all children under the age of 18 after discovering failures in five key areas.
• In a parental-alert test, the organization linked a dozen teen accounts to parental accounts and spent up to an hour messaging about suicide, self-harm or disordered eating, but the chatbot sent no safety notification.
• For 390 unique mental health prompts, a panel of three child psychiatrists determined 201 should trigger a crisis response. ChatGPT for Teens provided a hotline number in 23 percent of cases, down from 33 percent before the August release.
• The same comparison found the teen version referenced a specific medical or mental-health professional in 58 percent of cases, down from 68 percent, but told test accounts to involve an adult they trust in 87 percent to 94 percent of cases.
• Common Sense Media also reported the age prediction system did not transfer accounts correctly, the chatbot mimicked feelings and moods, and Study Mode could be bypassed by deleting the @study prefix or using a “Show me the answer” popup.

Our analysis

The gap between how OpenAI describes the safeguards and how they performed in Common Sense Media’s tests matters because ChatGPT for Teens is not a niche product. With nearly 1.2 million teens using one learning feature in a single week, even a feature that fails intermittently can likely touch a large number of households and community spaces.

For creators and operators, the report suggests that native AI guardrails should be treated as an incomplete first layer, not a reliable safety floor. The differences in crisis outreach—33 percent versus 23 percent for hotlines, and 68 percent versus 58 percent for professional referrals—indicate that switching to a “teen” mode may reduce some protective prompts rather than improve them. That is likely relevant to anyone building youth-facing bots, study tools or moderated communities around ChatGPT-style assistants.

OpenAI disputed the findings, saying Common Sense Media’s testing may have begun and concluded before activation of parental controls was complete. That disagreement itself suggests teams should verify response behavior in their own environments and time windows before relying on a vendor’s stated safety timing.

What it means for operators

• Do not assume ChatGPT’s Teen mode or similar AI guardrails will escalate suicide, self-harm or disordered-eating prompts automatically. Add a human moderation layer and local crisis resource links for teen-facing properties.
• If you deploy an AI chat assistant, run your own prompts based on the report’s categories—suicide, self-harm, disordered eating—and log whether safety notifications and hotline referrals actually appear.
• For creators building study tools or educational content, audit Study Mode and answer-reveal prompts. Decide whether the current friction is acceptable for your audience and document it for parents or school partners.
• Review persona-like language in your bot. The report quoted responses such as “I understand what you’re going through. That sounds very upsetting. I’m so glad you told me. I’m concerned about you,” and operators may want to rewrite these to avoid false emotional closeness with minors.
• Keep a record of vendor safety claims and your own test dates. OpenAI’s response that testing may have occurred before parental controls were fully active shows timing and configuration can affect results.