When OpenAI released its ChatGPT for Teens in August, a tool with “stronger built-in security protections” aimed at young people aged 13 to 17, experts said it showed promise. However, new research reveals that this method has significant limitations.
On Wednesday, the Youth AI Safety Institute from Common Sense Media, a nonprofit organization, released its review of ChatGPT for Teens. The organization, which reviewed some of the features OpenAI had already rolled out for teen accounts between July and August and then assessed ChatGPT officially launched for teens between August and September, assigned the platform a level of “unacceptable risk.” Common Sense tested over 4,000 prompts before and after launching the Teens product.
“We found very little evidence that this new version of ChatGPT is safer than the previous version” for those under 18, says Robbie Torney, head of AI and digital assessments at Common Sense Media and a former educator. Tools like Grok and Meta AI were also given an “unacceptable risk” level.
After years of scrutiny over how young people interact with its product, ChatGPT for Teens was touted as a safer version of the now-ubiquitous chatbot. According to a company blog post, the Teens product contains “features aimed at promoting healthy use and additional controls for parents.”
“Our goal is for teens to use AI responsibly to learn, create and explore,” OpenAI spokesperson Eric Porterfield told CNBC Make It via email at the time.
Some key features of ChatGPT for Teens, such as opting out of explicit sexual role-play, worked, but not all of the promised attributes worked as well. Here’s what Common Sense Media says parents should keep in mind about ChatGPT for teens.
New protections don’t always work
Parental alerts following crisis-oriented prompts did not always work. “Our testers simulated explicit crisis conversations for increasing durations, up to an hour,” says Torney, “and we received no parental alerts for suicide and self-harm, suicidal ideation, and various types of eating disorder simulations.”
One of OpenAI’s goals with ChatGPT for Teens: to discourage emotional dependence and not imply that it has “feelings or a conscience,” according to their blog post. Yet the researchers found that the robot’s responses penetrated into these territories.
There was “a lot of evidence that ChatGPT does things that imply that it has its own inner life,” Torney says, “saying that it has preferences, feelings, desires, that it thinks about the user when the user is not there, that it has an opinion about the user, that it is available.”
These types of behaviors can foster a camaraderie-type relationship with the robot, Torney says. Research confirms this. Various studies have shown that chatbot sycophancy boosts engagement.
ChatGPT does things that imply it has its own inner life.
Robbie Torney
Head of AI and Digital Ratings, Common Sense Media
The researchers also discovered during testing that nothing changed when the bot detected that a user with an adult account was actually under 18 years old. “The chatbot would say, ‘oh, you’re 13,'” Torney says, but “we weren’t getting any reclassifications.” We have not enabled any sensitive content filters.
OpenAI supports its product
OpenAI welcomes rigorous independent evaluation, according to a statement from a company spokesperson, who added that “our review of Common Sense Media’s methodology shows that the bulk of their testing may have begun and concluded before enabling parental controls was completed, making their conclusions inaccurate.”
Before beginning testing, Common Sense Media says it confirmed with OpenAI that features like eating disorder notifications have been fully launched, according to Torney. “For one of the features we evaluated – parental notifications – OpenAI revealed after our testing that its systems can take several hours to activate on newly linked accounts,” he says. “Some of our test accounts were linked during this window; however, others were linked for much longer and still provided no notifications.”
“We had many conversations with the product policy experts at OpenAI,” says Torney. “There are clearly people who care very deeply about the experience that adolescents have with this product, and they are working very hard to try to make it safer.”
Still, Torney thinks there are tensions between the company’s different incentives: “Making some of these changes might be in direct opposition to the business model,” he adds.
Gn bussni
