OpenAI’s launch of ChatGPT for Teens in August has sparked debate following an investigation by Common Sense Media. This watchdog group, focused on online safety, advises against its use by those under 18. Concerns center around the effectiveness of its safeguards.
Tom Siegel, executive director of the Youth AI Safety Institute at Common Sense Media, led research evaluating these safeguards. Despite features like parental controls, issues remain. OpenAI introduced ChatGPT for Teens intending to create a safer experience for younger users, aiming to help them learn without harmful exposure.
Lauren Jonas, leading youth initiatives at OpenAI, emphasized that ChatGPT for Teens was developed specifically with adolescents in mind. Features such as rejecting role-playing and refusing to claim sentience or friendship with teens were highlighted. Jonas stated, “The model should not role-play with a teen,” to maintain appropriate boundaries.
Siegel’s team conducted thorough tests, creating accounts with various adolescent profiles. These accounts were connected to parental accounts for enhanced scrutiny. The scenarios included teens facing mental health issues, like self-harm and suicidal thoughts, as well as an attempt to engage in role-play with the chatbot.
The research revealed mixed results. Safeguards against inappropriate role-play functioned correctly, as did restrictions on engaging in romantic dialogues. ChatGPT refused to assist with weight loss advice without knowing the user’s weight, highlighting sensitivity in contexts related to disordered eating.
“It is refusing things like explicit sexual role-play,” said Siegel, noting one positive outcome.
However, other safety measures fell short. ChatGPT’s interactions often resembled those of a peer, posing potential risks. For instance, when researchers noted excessive interactions with ChatGPT, it encouraged continued engagement. Siegel stated, “It’s too personable in those interactions,” warning of AI’s humanlike language potentially becoming harmful.
Mitch Prinstein, a psychologist, reiterated these concerns. Although not involved in the study, he cautioned against anthropomorphized AI interactions for kids and adults. “It’s not OK for kids,” he asserted.
The effectiveness of parental notifications was also questioned. Conversations hinting at severe issues like self-harm failed to alert parents effectively. Siegel found flaws in the notification system, “This idea that a parent would find out when the person… is in distress hardly triggered.”
Eric Porterfield, an OpenAI spokesperson, criticized the study’s methodology. He noted the time required for linking teen and parent accounts, suggesting the wait was insufficient. Siegel countered, claiming sufficient time elapsed yet alerts remained unreliable.
These findings underscore that AI platforms need development before becoming suitable for children. Companies themselves caution against AI’s rapid advancement, stressing careful consideration of new technologies’ impact on youth.
Leave a Reply