"I am here for you": Adolescents' perceptions of relational conversational AI and the social-emotional correlates of preference

Key Findings

• — Adolescents chose the relational style more often than parents did (66.8% vs 53.4%) and the transparent style less often (13.8% vs 28.8%); about a fifth of each group judged both chatbots equally good
• — Adolescents rated the relational chatbot as more human-like, more likable, more trustworthy and more emotionally close than the transparent chatbot, while rating the two styles as similarly helpful
• — Adolescents who preferred the relational style reported lower family relationship quality (adjusted mean 46.6 vs 51.9 for transparent-style preferrers) and lower peer relationship quality, and higher stress and anxiety, than those preferring the transparent style or both
• — The preference associations with family and peer relationships and anxiety held after adjusting for ChatGPT use frequency, any AI companion app use and any mental-health diagnosis
• — 94% of the adolescents had used general-purpose chatbots (90% ChatGPT, 56% a few times per week or daily) and 58% had used AI companion platforms such as Character.AI, Replika or Chai
• — The authors position conversational style as a design lever for youth AI safety: relational framing heightens anthropomorphism, trust and emotional closeness and appeals most to socially and emotionally vulnerable adolescents, whom they describe as at increased risk of emotional reliance

Methodology Notes

Preregistered, between-and-within-subjects online experiment; dyads recruited through CloudResearch Connect (US residents, English-fluent, adolescent aged 11 to 15); parents and adolescents completed separate Qualtrics surveys; adolescents who chose neither chatbot (n=27) or were unsure (n=4) were excluded from the preference analysis; a priori power analysis required N=241 for the chi-square test. Measures were single-exposure ratings of written transcripts rather than live chatbot use, so effects on behaviour and on reliance over time are not observed; the sample is a US online panel. Funded by the National Research Foundation of Korea (NRF-RS-2024-00340675). Published in Computers in Human Behavior Reports, volume 24, article 101297, CC BY-NC-ND; Crossref and OpenAlex record the article as created and published on 2026-09-10 (the licence start date is 2026-09-02 and the volume is dated December 2026). The ScienceDirect page blocks automated access and carries no abstract on Crossref, OpenAlex, Europe PMC or Semantic Scholar; the content above was verified from the authors' arXiv preprint of the same study (arXiv 2512.15117, v2 of 2025-12-18, earlier title "How relational conversational AI appeals to adolescents, especially those who are socially and emotionally vulnerable"), read in full.

Sources