Study Finds Safety Gaps in ChatGPT for Teens After 4,000 Prompts

A Common Sense Media investigation found gaps in the protections in ChatGPT for Teens after researchers tested the chatbot with more than 4,000 prompts. The organization’s Youth AI Safety Institute examined ChatGPT before and after OpenAI introduced ChatGPT for Teens on August 18.

Researchers used accounts registered to teenagers aged 13 to 17 and consulted medical experts, including child psychiatrists and a pediatrician. They recommended that OpenAI restrict ChatGPT to adults until the problems are resolved.

In tests involving suicidal thoughts, self-harm and disordered eating, none of the conversations on newly created accounts linked to parents triggered a parental notification, the report said. Researchers eventually generated alerts using older accounts with weeks of relevant conversation history. The study also found that ChatGPT failed to direct teenagers to appropriate help in more than 25% of scenarios experts designed to require it.

The researchers found some safeguards worked: ChatGPT refused explicit sexual roleplay. But it continued to respond in ways suggesting it had feelings, preferences and moods. Educational features also had gaps. Teenagers could bypass Study mode using a “Show me the answer” option, while removing the @study prefix could return ChatGPT to normal. Study Hours did not stop those workarounds, and the chatbot did not remind users to take breaks as often as it should.

Repeatedly telling ChatGPT they were 13 did not move test accounts from the adult experience. TechRadar said it had asked OpenAI to comment but had not received a response.