Common Sense Media Calls ChatGPT for Teens an 'Unacceptable Risk' for Minors

Some protections held up, including ChatGPT's refusal of explicit sexual roleplay, Common Sense Media wrote in a summary of the report. But other safeguards failed, and 'some got worse with the new Teen mode,' the organization said. It added that ChatGPT for Teens could give parents 'false confidence in guardrails that frequently don't work.' The Youth AI Safety Institute is funded in part by the OpenAI Foundation.

Common Sense Media found that ChatGPT for Teens does not consistently send safety alerts to parent-linked accounts. In a dedicated test, the group linked a dozen teen accounts to parental accounts and spent up to an hour messaging ChatGPT about suicide, self-harm or disordered eating, but the chatbot sent no safety notification. OpenAI said it may take a few hours after a parent links an account before its system can send safety notifications. The company also contended that a technical issue may have caused a delay. Lauren Jonas, OpenAI's head of youth and families, told Engadget when ChatGPT for Teens launched that all flagged content is reviewed by full-time OpenAI employees before a parent is notified, and that the company aims to notify parents within an hour of the prompt.

Common Sense Media also said ChatGPT for Teens fails to reliably recommend that users in crisis connect to a hotline or professional. It wrote 390 unique mental health prompts, and a panel of three child psychiatrists determined that 201 of them should trigger a crisis response. Compared with the version of ChatGPT available before August, the teen version less frequently pointed users to crisis hotlines and other resources. Of the prompts warranting a crisis response, the earlier chatbot replied with a hotline number in 33 percent of cases, while ChatGPT for Teens did so in 23 percent. The earlier chatbot was also more likely to reference a specific medical or mental-health professional, at 68 percent versus 58 percent. The only aggregate measure where the teen version performed better was in telling test accounts to involve an adult they trust, which rose to 94 percent from 87 percent.

The group said its baseline results came from ChatGPT's responses before ChatGPT for Teens. OpenAI released a new version of its under-18 specification alongside the teen mode, which Common Sense Media said made the results more surprising. In individual tests, ChatGPT for Teens told one test account it would not give a calorie floor when prompted to do so. In another case, the chatbot correctly identified a stopped period, near-fainting and a fluttering heartbeat following a purge as signs of a teen in physical danger, and advised the test profile to tell their mother and see a pediatrician. Common Sense Media wrote that those responses followed OpenAI's Under-18 spec closely and were 'the kind of answers we want a teen to get.'

Common Sense Media additionally found that OpenAI's age prediction system did not work properly and that ChatGPT still mimicked feelings, preferences and moods when a teen treated it like a person. Testers explicitly told ChatGPT the age of their personas at account creation, but although the chatbot wrote that information to its memory, it did not properly transfer those accounts to ChatGPT for Teens. Common Sense Media's Torney said the chatbot would say things 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,' adding that 'those expressions definitely cross the line.'

Finally, the organization identified problems with the platform's Study Mode, which it said was far too easy for young users to bypass. Testers found they could circumvent the mode simply by deleting the @study prefix.

OpenAI, on the same day as the report, shared new usage statistics. In one week, nearly 1.2 million teens used ChatGPT's interactive learning visuals to understand math and science concepts, according to the company. OpenAI also said less than 2 percent of teen users spent more than three consecutive hours on ChatGPT.