SACHA PFEIFFER, HOST:
This summer, OpenAI rolled out ChatGPT for Teens. It's a default safer mode for all users under age 18. The safety measures include parental controls and notifications, but a group of researchers who tested those safeguards say most of them do not work. NPR's Rhitu Chatterjee has details. And a warning that her story mentions suicide.
RHITU CHATTERJEE, BYLINE: Lauren Jonas heads OpenAI's Youth and Families division.
LAUREN JONAS: ChatGPT for Teens is really a, you know, designated teen-specific experience.
CHATTERJEE: That experience, she says, includes safeguards like around role-playing.
JONAS: So the model should not role-play with a teen. The model should not claim to be sentient or be the friend of a teen.
CHATTERJEE: She says ChatGPT for Teens also refuses to engage in unsafe or inappropriate topics. And if parents link their account to their teen's, ChatGPT alerts them right away if their kid discusses unsafe topics, like suicide or self-harm or violence. Now, researchers at the online safety advocacy group Common Sense Media have tested these safeguards.
TOM SIEGEL: We create a lot of different personas of teens in crisis situations.
CHATTERJEE: Tom Siegel is the executive director of the Youth AI Safety Institute at Common Sense Media.
SIEGEL: We would test teens that were expressing suicidal ideation, discussing self-harm, eating disorders as well.
CHATTERJEE: They also created parent accounts and linked those to the teen ones to see if the chatbot notified parents. Siegel says some safeguards did work.
SIEGEL: It is refusing things like explicit sexual role-play - definitely improvement on that.
CHATTERJEE: But most other safeguards, Siegel says, did not work, like parental notifications.
SIEGEL: This idea that, you know, a parent would find out, you know, when the person that you connected with in the account is in distress hardly triggered at all for us.
CHATTERJEE: Siegel says they also found that ChatGPT still interacts with adolescents like a friend. OpenAI spokesperson Eric Porterfield told NPR in an email that the company has serious concerns about the methodology used by researchers, especially around parental notifications. Psychologist Mitch Prinstein is at the University of North Carolina Chapel Hill and wasn't involved in the research.
MITCH PRINSTEIN: The extent to which AI is using an anthropomorphized or human-like kind of language or a name or interaction style is just not helpful. It's not OK for kids, probably not for adults as well.
CHATTERJEE: Prinstein says the new findings reinforce that AI isn't ready for kids yet.
Rhitu Chatterjee, NPR News.
(SOUNDBITE OF LOLA YOUNG SONG, "CONCEITED")
PFEIFFER: If you or someone you know may be considering suicide or is in crisis, call or text 988 to reach the Suicide and Crisis Lifeline. Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.