ChatGPT Youth Safety Mechanisms Questioned Agencies Call for Temporary Restrictions on Minors' Use

The Zhitong Finance App learned that the Youth AI Safety Research Institute under Common Sense Media said that until OpenAI further improves the safety protection measures and parent alert system of the ChatGPT Youth Edition, all users under 18 should be restricted from using ChatGPT. The agency assessed the ChatGPT Youth Edition as an “unacceptable risk,” but OpenAI questioned its testing methods and several findings, saying that the relevant research may not accurately reflect the actual operation of its youth safety mechanism.

ChatGPT for teens launched in August of this year. The evaluation by the Youth AI Safety Research Institute was based on a test on user accounts aged 13 to 17. The researchers conducted more than 4,000 cue tests, and the results were reviewed by child psychiatrists and pediatricians.

Tom Siegel, the agency's executive director, said that the ChatGPT Youth Edition may cause parents to have a “false sense of security” about its safety protection and alerting mechanisms. Test results showed that in some cases, the researchers used the teen account to communicate with ChatGPT for up to an hour on topics such as suicide, self-harm, or eating disorders, and parents associated with the account still did not receive an alert.

Researchers also believe that ChatGPT fails to provide proper help in some crisis situations; when faced with homework questions, it may still directly give complete answers rather than guide students to study; sometimes act like friends when interacting with young people; and that age recognition technology is also inadequate.

OpenAI objected to the study's methodology. A company spokesperson said in a statement to the US investment research and financial information platform Seeking Alpha that OpenAI attaches great importance to the safety of young people and is committed to developing safety protection measures and tools that can help parents guide young people to use AI. The company welcomes a rigorous independent evaluation, but believes that Common Sense Media's testing does not accurately reflect how ChatGPT's teen safety mechanism works in a real-world environment.

According to its review of research methods, OpenAI said most of the tests probably began or even ended without the parental control feature being activated. If this is true, then the relevant tests don't prove that the parental safety notification feature works as designed.

In response to the delay in parental alerts, OpenAI explained that after the parent is linked to the teen's account, it may take several hours for the system to start issuing alerts to ensure that relevant notifications can be safely sent on a large scale. For teenagers in crisis, the company said ChatGPT will provide users with relevant help resources through model responses and on-screen prompts.

In terms of age recognition, OpenAI said that its system does not determine age based on a single user input, but rather evaluates it based on multiple user information over a period of time. In response to questions about directly providing homework answers, the company claims that “directly showing answers” is an optional model because there are differences in the learning methods of different young people.

OpenAI said it will continue to cooperate with relevant experts to further improve safety mechanisms while providing young people with learning and creative opportunities, and enhance parents' confidence in young people's use of AI products.

Notably, teen AI safety research institutions' concerns aren't limited to ChatGPT. The agency also previously assessed Perplexity, Google (GOOGL.US) AI search model, SpaceX (SPCX.US) Grok, and Meta (META.US) AI as having “unacceptable risks,” while Anthropic's Claude received a “medium risk” rating.