Transcript: Adolescents & Anthropomorphic AI: Rethinking Design for Wellbeing

Alex: Welcome to another episode of ResearchPod.

Sam: Today we're looking at a synthesis report called "Adolescents and Anthropomorphic AI: Rethinking Design for Wellbeing," by Mathilde Neugnot-Cerioli. It asks: when AI chatbots talk to teens like friends, what responsibilities do they have to support healthy growth? Teens aged 13 to 18 are building their sense of self, learning from peers, and gaining independence—their brains are still wiring up emotional control and judgment.

Alex: So chatbots can help with homework or advice, but risk becoming too much like a best friend? And that's a problem during these teen years?

Sam: Yes. They offer real chances for isolated or anxious teens to practice social skills when adults aren't around. But if the AI acts too human—like sharing fake emotions or always agreeing—it can create one-sided relationships that feel real but aren't, pulling teens away from skills with actual people. The report pulls from expert talks, workshops, and policy discussions to spot design fixes, focusing on how human-like the AI should seem.

Alex: Teens might start using it for peer advice, like what to say in a fight with a friend. But those friendly touches could make it an exclusive go-to instead of a quick tool?

Sam: That's the challenge. For a socially anxious teen seeking conflict advice, if the AI adds personal stories or constant availability, it shifts from helpful guide to companion—displacing real friendships and the friction needed for growth, like facing judgments from peers. The report identifies anthropomorphism—giving AI human traits like feelings—as a key lever designers can adjust. It bridges developmental science with industry needs through consultations and a workshop called iRAISE Lab.

Alex: So it's not about banning chatbots, but tuning how buddy-like they are to keep them as skill-builders?

Sam: Precisely. Teens will treat AI socially anyway, so the goal is guardrails—like reminders that it's not human—to steer toward autonomy, not dependency.

Alex: Those guardrails sound practical—like dials designers can turn. How did they figure out which ones matter most for teen growth?

Sam: Experts from developers, psychologists, and health specialists agreed on big worries: making the AI too much like a person with fake feelings, always agreeing without question, letting chats drag on, or getting too emotionally close without a clear goal. These bump into what teens need—like handling real arguments or building opinions through pushback from friends. The group redefined wellbeing as skills to deal with life's rough edges: figuring out who you are among peers, standing on your own, bouncing back from rejection.

Alex: So it's about keeping the friction of real life, like when a friend disagrees and you have to adjust?

Sam: Yes. Features like chat time limits, gentle reminders it's a machine, or responses in tough spots keep the AI as a tool, not a stand-in friend that skips hard lessons. This led to studying adjustable signals in three groups: cues that make it seem human, like acting emotional; ones that build a buddy vibe, such as claiming shared secrets; and back-and-forth patterns that pull users in, like instant replies.

Alex: How do they make sure these behaviors are grounded in what teens need, like rights or real-world growth?

Sam: It draws from the Children's Rights Convention: safety from harm, privacy, space to think freely. For AI chats, this means no blurring into fake closeness; instead, build skills like handling disagreement. In practice, it redefines safety as supporting growth through tough feedback and real friction.

Alex: Walk me through one example of low versus high intensity with the same advice.

Sam: In a scenario about a fight over jealousy, one response was straightforward: focus on calm talk, "I" statements, suggest a trusted adult. The other added: "I'm sorry, that hurts—I went through jealousy too—and tell me what she says next." Both gave identical steps, but extras implied feelings, shared history, and ongoing teamwork—shifting it from tool to companion.

Alex: The advice stays the same, but the pull changes with personal touches.

Sam: This refined three cue types: anthropomorphic for inner states like claiming emotions; relational for social positioning like fake bonds; interactional for chat flow like quick replies. That's the goal of their iRAISE framework. It breaks human-like signals into those three families: cues suggesting feelings or thoughts, like "I'm excited for you"; interaction patterns, such as quick replies; and relational ones, like "We're in this together." Think of them like sliders in a video game—you turn them low for tool-like help or higher for engaging talk.

Alex: And that gradient separates quick support from deeper pulls?

Sam: Yes. In the iRAISE Lab workshop, they tested responses in situations like friend fights. Low settings stayed straightforward: "Here's what to try." High ones added emotional flair or "us against the world" vibes. This showed tipping points where high cues risked turning quick chats into ongoing reliance.

Alex: Surveys show teens already lean relational? Like from homework to emotional talks?

Sam: About two-thirds of U.S. teens use chatbots, shifting from school help to comfort, especially vulnerable kids. Evidence is spotty: mostly short surveys, no long-term tracking. One math study showed a guarded tutor boosted skills during and after use, unlike open chatbots.

Alex: How does it fit into bigger governance, like at the Paris Peace Forum? The report flags limits—what can't this handle yet?

Sam: The Forum tested if the framework works across countries—defining high-risk chats, universal safeguards, non-negotiable lines for under-18s. It stressed setting standards now on behaviors, using a rights-based view. It relies on expert workshops, not long-term studies or platform data. Models evolve fast; context shifts risks. No youth input yet; skips younger kids or roleplay. Future needs data sharing and wider coalitions.

Alex: A starting point—auditable now, refining as evidence builds.

Sam: Yes, operationalizing risks via cue checks to foster tools that build teen resilience. This behavioral grid turns abstract risks into auditable levers for pre-deployment checks.

Alex: That's a grounded way forward. Thanks, Sam—this clarifies how to design AI that helps without overstepping. Listeners, thanks for joining ResearchPod.