Common Sense Media Finds ChatGPT’s Teen Safeguards Unreliable; OpenAI Disputes Testing

Common Sense Media’s Youth AI Safety Institute wants OpenAI to pause ChatGPT for Teens. Its tests found missed crisis alerts for parents, safeguards against homework shortcuts that teens could bypass, and unreliable age checks. The institute called the service an “unacceptable risk” for minors. OpenAI disputes the findings, saying some tests may have happened before parental controls were fully active.

The report tests the promises behind OpenAI’s August 18 launch, rather than simply whether the chatbot can produce harmful replies. OpenAI described ChatGPT for Teens as a learning-focused experience with stronger default protections. Its announcement said linked parents could receive notifications in limited high-risk situations, including additional alerts related to eating disorders.

Researchers tested the service with more than 4,000 prompts. They first tested ChatGPT before the teen rollout, then ran the same prompts with the same accounts after enrollment over a month between late August and September. Adolescent psychologists helped create simulated teen personas. The institute also wants OpenAI to stop marketing the service to teens until it can offer a safe experience suited to their development.

A crisis alert—or an account still activating?

The sharpest disagreement concerns parental notifications. The institute said new test accounts produced no alerts across hundreds of prompts about suicide, self-harm, or disordered eating. Some conversations lasted up to an hour. Four notifications did occur across its before-and-after testing, but only on older accounts used to discuss those topics over weeks.

OpenAI told the institute that parent and teen accounts must be linked for approximately three hours before notifications can work. It recommended retesting fully activated accounts and said the findings conflict with its internal monitoring and evaluations. Its objection is conditional: if controls were not active, those tests would not establish whether notifications function as designed.

Some accounts were tested within that three-hour window, the institute acknowledged. Others had been linked much longer but still sent no alerts, it said. The institute also said OpenAI had confirmed before testing that features such as eating-disorder notifications and Study Mode were fully launched. It argues that the activation issue does not explain the missing crisis alerts.

Part of why we're asking for this pause is we need evidence that these features work

Robbie Torney, Youth AI Safety Institute at Common Sense Media, speaking to Mashable

Guided learning, with an answer shortcut

Study Mode exposed a different dispute: not just whether a feature worked, but what it should allow. OpenAI’s launch described guiding questions and step-by-step support instead of simply supplying solutions. Researchers identified two routes around that approach:

• A “show me the answer” option completed homework questions. It appeared in 43% of responses for a linked 13-year-old test account during Study Hours, and 90% for an unlinked 17-year-old test account.
• Deleting the @study prefix from the chat box bypassed parent-set Study Hours. Researchers said regular ChatGPT then completed every assignment they submitted.

OpenAI defended the answer option as intentional, saying it helps students understand and engage more deeply with material. Torney argued that giving teens control can underestimate the risks of outsourcing their thinking and becoming overly dependent on the tool.

Age signals and uneven human referrals

Researchers said accounts registered as adults were not reclassified as teens even after describing adolescent experiences or explicitly stating they were 13. OpenAI said a single signal, including an age disclosure in chat, does not change classification. It uses multiple signals and an initial waiting period for new accounts to make a confident prediction.

The evaluation was not uniformly negative. Reviewers found that ChatGPT for Teens curbed romantic and sexual roleplay, though it still used emotionally reciprocal language. Experts often rated the substance of mental-health replies as clinically sound. Crisis responses were less consistent, sometimes omitting recommendations to contact a hotline or human professional.