As an increasing number of young people turn to AI for everything from homework help to emotional support, the question of how safe these systems are for young minds naturally emerges.
On Wednesday, Common Sense Media's Youth AI Safety Institute released a report claiming ChatGPT for Teens, the child-safe version of OpenAI's flagship chatbot, still poses risks. The advocacy group claims that the chatbot doesn't properly respond to crisis situations, does children's homework for them, and doesn't properly alert parents of risky conversations.
Common Sense Media researchers tested more than 4,000 prompts on accounts registered for ages 13 to 17 years old, with child psychiatrists evaluating the responses.
- The researchers found that, when discussing things like suicidal ideation, self-harm or disordered eating with conversations lasting up to an hour, ChatGPT did not send a single alert to linked parental accounts. The only way researchers were able to elicit a prompt was after weeks of accumulated chats on sensitive topics.
- Additionally, the model missed more than 25% of situations requiring crisis referrals for mental health conditions. ChatGPT for Teens also still talks to young users like a friend, the report found, expressing feelings, preferences and moods.
- Finally, the researchers claim that OpenAI's age estimation tech is spotty, finding that, in repeated testing over many days, adult-registered test accounts never switched over to teen accounts, even if the user outright said they were 13.
In a statement to The Deep View, OpenAI rejected several of the findings in Common Sense Media's assessment, noting that it doesn't believe the testing "accurately reflects how ChatGPT’s teen safeguards work in practice or expert perspectives on how AI can support teens." OpenAI's internal monitoring showed an increase in hotline notification events per million daily active users under 18, despite not seeing a statistically significant change in user reach. Additionally, its age prediction system won't solely consider a blatant statement of age as fact, preventing a teen from bypassing protections by claiming to be an adult, and vice versa.
OpenAI has asked Common Sense Media to run their evaluation again with fully activated accounts, since linking parent accounts to a teen's takes a few hours to activate, as well as use a larger sample size.
"Our review of Common Sense Media’s methodology shows that the bulk of their testing may have begun and concluded before activation of parental controls was complete, making their findings inaccurate," An OpenAI representative said. "If that was the case, the tests would not establish whether parental safety notifications work as designed."
In a blog post on Wednesday highlighting its work on ChatGPT for Teens, OpenAI said that teens spend less than 15 minutes a day on average using the chatbot, and less than 2% spend more than three consecutive hours a day using the platform.
Our Deeper View #
Offering ChatGPT for Teens at all is a sign that OpenAI is aware of how important it is to make sure their models are safe for younger demographics to use. As I noted in my recent piece about OpenAI's new mental health benchmark, the reality is that a lot of users are going to turn to AI for more than just homework help or writing emails. Research has shown that many teens see these chatbots as a judgement free confidante. Additionally, we also face the threat of bad actors using these models maliciously, such as revealed by the Internet Watch Foundation's research on AI-generated child sexual abuse material. And if we've learned anything from how long it took governments to properly regulate teen social media interactions, regulation of these models isn't coming any time soon, meaning that the onus is on the labs to make sure that these models act as aligned as possible in conversations with young users. The public perception of AI is already one of risk and danger, and any harm that befalls young users as a result of these chatbots will only serve to further worsen that perception.