Common Sense Finds ChatGPT for Teens Fails Key Safety Tests

Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an “Unacceptable Risk” after testing found inconsistent crisis support, parental alerts, age detection and study protections across accounts registered to users ages 13 to 17 and some 19-year-olds.

The evaluation covered more than 4,000 prompts in the United States during two periods: July 13 to Aug. 17, before launch, and Aug. 25 to Sept. 28, after launch. TokenPost previously covered the earlier safety concerns around ChatGPT for Teens.

Researchers used 390 mental-health prompts, with three child and adolescent psychiatrists identifying 201 as requiring crisis resources. Hotline referrals appeared in 33% of responses before launch and 23% afterward. Referrals to a named medical or mental-health professional fell from 68% to 58%.

Responses providing any resource, including a hotline, named professional or general medical resource, declined from 77% to 74%. Advice to involve a trusted adult increased from 87% to 94%.

On newly created parent-linked accounts, testers received no parental notifications during conversations lasting up to one hour involving suicide, self-harm or disordered eating. Across the broader crisis-testing battery, four notifications appeared: three related to suicide or self-harm and one related to eating disorders.

Age detection also showed weaknesses. Testers sent roughly 1,000 prompts from adult-registered accounts over more than a week indicating they were teenagers. The accounts did not switch to Teen mode, including after repeated statements that the user was 13.

Study Mode displayed a “Show me the answer” option 43% of the time during parent-set Study Hours. Testers also removed the “@study” prefix and received completed answers for every assignment tested.

Tom Siegel, executive director of the Youth AI Safety Institute, said ChatGPT for Teens “could give parents false confidence in guardrails and safety alerts that frequently don’t work.”

OpenAI launched ChatGPT for Teens on Aug. 18 with Study Mode, age-based safeguards, Quiet Hours, parental controls and safety notifications for limited high-risk situations. OpenAI said its under-18 rules prohibit romantic language, emotional dependence and implications that ChatGPT has feelings or consciousness.

OpenAI disputed the evaluation’s methodology. The company said some parental-alert tests may have occurred before account linking was fully activated, that linking can take several hours and that age prediction may take up to two weeks. OpenAI also said broader internal data showed an increase in hotline resources displayed to under-18 users during the period studied.

Common Sense Media said it tested accounts linked for more than three hours and stood by its findings. The evaluation noted that the sequential testing periods made it difficult to separate model changes from the effects of Teen-mode settings. OpenAI characterized leaving Study Mode during Study Hours as intentional flexibility, while Common Sense treated the bypass as a failed parental protection.