When OpenAI released ChatGPT for Teens in August, a tool aimed at 13- to 17-year-olds “with stronger built-in safeguards,” experts said it showed promise. However, new research reveals that this has important limitations.
On Wednesday, media ratings nonprofit Common Sense Media’s Youth AI Safety Institute published a review of ChatGPT for teens. The group examined some of the features OpenAI had already rolled out for teen accounts in July and August, and conducted an evaluation of ChatGPT for teens, which was officially released in August and September, giving the platform an “unacceptable risk” level. Common Sense tested over 4,000 prompts before and after the launch of our Teens product.
“We found little evidence that this new version of ChatGPT is any safer for children under 18 than previous versions,” said Robbie Toney, head of AI and digital assessment at Common Sense Media and a former educator. Tools like Grok and Meta AI have also reached “unacceptable risk” levels.
After years of research into how young people interact with its products, ChatGPT for Teens was touted as a safer version of the now ubiquitous chatbots. According to a company blog post, the Teens product includes “features that promote healthy usage 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.
While some key features of ChatGPT for Teens worked, such as rejecting explicit sexual role-play, not all of the promised attributes worked as well. Common Sense Media says parents should keep in mind ChatGPT for Teens.
New protections don’t always work
Crisis-oriented prompts followed by parental warnings did not always work. “Our testers increased the amount of time they simulated overt crisis conversations, up to an hour, and received no warnings from parents about simulating suicide, self-harm, suicidal ideation, or different types of eating disorders,” Toney says.
One of OpenAI’s goals with ChatGPT for Teens is to limit emotional dependence and not imply that it has “feelings or consciousness,” according to their blog post. Still, the researchers found that the bot’s responses penetrated these areas.
Torney said, “There was a lot of evidence that ChatGPT was doing things that suggested it had an inner world of its own: it had preferences, emotions, and desires; it was thinking about you even when you weren’t there; it had opinions about you; it was available to you.”
These types of behaviors can foster a peer-like relationship with the bot, Tawney says. Research supports this. Various studies have shown that chatbot flattery increases engagement.
ChatGPT does things that suggest it has an internal life of its own.
robbie toney
Head of AI and Digital Evaluation, Common Sense Media
In their tests, the researchers also found that when the bot detected that a user with an adult account was actually under 18, nothing changed. “The chatbot was like, ‘Oh, you’re 13 years old,'” Toney says, but “there was no reclassification. None of the sensitive content filters were turned on.”
OpenAI stands behind our products
OpenAI welcomes the rigorous independent evaluation, according to a statement from a company spokesperson, adding: “Our review of Common Sense Media’s methodology found that the majority of the tests may have started and ended before parental controls were enabled, resulting in inaccurate results.”
Tawney said Common Sense Media confirmed with OpenAI that features such as eating disorder notifications were fully implemented before testing began. “For parental notifications, one of the features we evaluated, OpenAI revealed after testing that it could take several hours for the system to become active on newly linked accounts,” he says. “While some of the test accounts were linked within that time period, other accounts were linked for quite some time and were not provided with notification.”
“We had a number of conversations with OpenAI’s product policy experts,” Torney says. “It’s clear there are people who are very concerned about what teens will experience with this product, and they’re working hard to make it safer.”
Still, Toney believes there is tension between the company’s various incentives, adding that “making some of these changes could go directly against the business model.”
