Kim, Xie and Yang: relational AI language appeals to adolescents, but no observed displacement

Pilyoung Kim, Yun Xie and Sujin Yang, Computers in Human Behavior Reports, available 10 September 2026; author preprint with methods and tables. Read alongside child-development question.

Identification and result. Preregistered U.S. online study, 284 parent–child dyads, children 11–15. Both read two matched written AI responses to an ordinary peer exclusion problem; order randomized, relational style used first-person warmth/commitment and transparent style stated nonhuman limits and gave similar practical advice. Among substantive preference choices (excludes 31 youth and 19 parents who said neither or unsure), 66.8% of youth selected relational versus 53.4% of parents. Youth rated relational language more human-like, likable, trustworthy and emotionally close; helpfulness difference not significant (p=.082). Youth who preferred relational language reported worse pre-existing family/peer relationship quality and more stress/anxiety; those associations are cross-sectional, not evidence chatbot use made them distressed. See preprint pp. 5–7, 14–16, 21.

Causal boundary. The experimentally moved endpoint is a hypothetical style judgment, not actual companion interaction, repeated attachment, human-friend contact, crisis response or teen mental health change. The transparent alternative was also more informational: experiment does not isolate the one-sentence AI disclosure from warmth and framing. Balanced response content and similar-rated helpfulness are promising design clues, not a demonstrated long-run safe setting. Test in a live multi-week randomized adolescent setting with an age-appropriate warm but bounded arm, scripted crisis/human handoff, passively recorded human contact, validated loneliness and parent/peer relationship outcomes, and safe consent/withdrawal. Current preference correlation could mean lonely youth seek low-judgment support (benefit) as readily as it could indicate risk of displacement.