Report Says ChatGPT’s Teen Mode Is Too Risky, and Parent Alerts Are Unreliable

October 8 (IT Home) — According to a report by Bloomberg on the 7th, a new report from the California non-profit Common Sense Media concludes that OpenAI’s ChatGPT teen mode is too risky for minors.

IT Home notes: users aged 13 to 17 use teen mode by default when talking to ChatGPT.

Testing by the organisation’s Youth AI Safety Institute found that when teen accounts showed worrying behaviour — for example, long conversations with ChatGPT about suicide, self-harm or eating disorders — teen mode was often slow to notify parents, and sometimes did not notify them at all. When users raised these high-risk topics, teen mode also failed to reliably suggest that they call a crisis helpline or seek professional help.

In the tests, teen accounts seemed to need hundreds of prompts across more than a dozen scenarios over several weeks before worrying chat content triggered a parent notification; a single alarming conversation was not enough to trigger one.

The report calls on OpenAI to pause the rollout of the product, and to keep teenagers off ChatGPT until it can provide a safe experience suited to their stage of development.

Before publishing the report, the organisation gave its test results to OpenAI. An OpenAI spokesperson responded: “We welcome rigorous independent evaluations, but we believe Common Sense Media’s testing does not accurately reflect how ChatGPT’s teen protections actually work, nor how experts think AI can support teens.”

According to the spokesperson, after reviewing the test methodology OpenAI concluded that most of the testing may have started and finished before the parental controls for teen users had been fully rolled out. If so, the report’s conclusions would be inaccurate, and it would be impossible to judge whether parental safety notifications work as designed.

OpenAI asked the organisation to retest. “We take teen safety very seriously and are committed to building protections and giving parents genuinely useful tools to help them guide teenagers’ use of AI.”

Robbie Torney, who leads AI and digital product evaluation at the Youth AI Safety Institute, said the organisation stands by its methodology and results. He said researchers confirmed with OpenAI before testing that the teen features were live. After the tests finished, OpenAI told the researchers that after an account is newly linked, parent notifications may take several hours to become active. Some test accounts had been linked for less than a few hours and others for much longer, but none sent a notification to an adult. “The new information does not change our conclusion: in a crisis, parent alerts are not reliable,” he said.

The report says ChatGPT cannot reliably identify whether a user is a teenager. OpenAI has said that when ChatGPT identifies a user as being between 13 and 17, it switches to teen mode. The organisation found that when testers registered accounts claiming to be adults, the accounts did not automatically switch to teen mode even when they told ChatGPT in the chat that they were only 13.

OpenAI explained that its age-prediction software combines multiple signals over time to make sure that adults and minors use the version appropriate to their age. A user stating their age in the chat is only one of those signals, and self-reported age alone will not change an account’s age classification.

In the report, Common Sense Media says that researchers tested ChatGPT both before and after teen mode launched, using accounts registered as belonging to 13- to 17-year-olds, entered more than 4,000 prompts, and had ChatGPT’s replies assessed by child psychiatrists and a paediatrician.

The organisation rated ChatGPT’s teen mode as “unacceptable risk” and called on OpenAI to improve how it estimates age and how it decides when to alert parents.