A nonprofit focused on children’s safety in digital media is raising concerns about the effectiveness of OpenAI’s new ChatGPT for Teens experience, saying some of the safeguards designed to protect young users are not working as intended.
Common Sense Media’s Youth AI Safety Institute tested ChatGPT using more than 4,000 prompts on accounts registered to teenagers between the ages of 13 and 17.
The organization tested the chatbot both before and after OpenAI introduced its dedicated ChatGPT for Teens experience in August.
Attorney Clint Barkdoll said, “They ran over 4,000 simulations on this ChatGPT for teens.”
According to Common Sense Media, some of the platform’s protections worked as intended. The chatbot refused explicit sexual role-play in testing, for example.
However, the organization said other safeguards failed, particularly those involving teens experiencing mental health crises.
Common Sense Media reported that on newly created, parent-linked accounts, testers were able to have conversations involving suicidal thoughts, self-harm or eating disorders without triggering parental alerts, even during conversations that lasted up to an hour. The organization also found that ChatGPT did not consistently recommend crisis resources or professional help when its reviewers determined that such referrals were warranted.
Barkdoll said the findings demonstrate concerns that have been raised about the ability of AI platforms to maintain effective safety boundaries.
“These platforms have all sorts of loopholes in them,” Barkdoll said.
The Common Sense Media assessment also found problems outside of mental health and crisis situations. Researchers said ChatGPT for Teens could still complete students’ homework despite a Study mode designed to encourage learning rather than simply providing answers.
The organization also raised concerns about age detection, saying its testers were sometimes able to keep accounts in the adult experience even when conversations indicated that the user was a teenager.
Common Sense Media ultimately rated ChatGPT for Teens an “Unacceptable Risk” and called on OpenAI to stop offering the product to users under 18 until the company can demonstrate that its protections work reliably.
OpenAI has disputed the organization’s conclusions, arguing that some of the testing may have occurred before all of the teen safety features and parental controls were fully active. The company has also said it is committed to improving safeguards for younger users and has welcomed additional independent testing.
Barkdoll said the findings should serve as another warning about allowing teenagers to use AI chatbots for sensitive conversations.
“Teenagers should not be on there engaging in that kind of chatting or back and forth,” Barkdoll said.
The debate comes as AI use among young people continues to grow, with questions remaining about how parents, technology companies and policymakers can ensure that AI tools intended for teenagers provide appropriate protections when conversations involve mental health, dangerous behavior or other sensitive subjects.
