How US Youth Use AI Chatbots: Conversation Patterns From Naturalistic Keystroke Observations

Key Findings

• — 62% of user-app-days involved using AI as a tool (26,343/42,355); 15% involved violence (6,460), 15% sexual role-play (6,393), 9% general role-play (3,707), 8% romantic role-play (3,538) and 6% friend-like interaction (2,624)
• — Violence and sexual or romantic role-play were relatively more common among younger users; tool use was more common among older adolescents
• — Companionship or role-play apps showed more sexual, romantic, general role-play and violent content, while general-purpose apps were mostly used as tools; ChatGPT appeared across all themes because of its high overall use
• — Word counts were highest on user-app-days involving violence, then sexual role-play, romantic role-play and emotional support
• — Violence, sexual role-play and romantic role-play frequently co-occurred within a user-app-day, including instances the authors describe as suggestive of violent sexual role-play
• — Engagement was highly variable: median 4 days of use (IQR 1-11) and a median of 6 messages per app-day

Methodology Notes

Deidentified keystroke data from a commercial parental-monitoring app; the sample is limited to children whose guardians installed Aura and who opted into the Aura keyboard and used GenAI apps, so it is not representative of US youth. Study period January-September 2025 (273 days, as stated; day-level bounds not given). Themes coded by an LLM with human review. WCG IRB waiver of consent (WCG20243405). The last author is affiliated with Aura; some authors report paid consulting or expert-witness roles in US social media litigation. The full text notes a prior paper on the same sample (prevalence and frequency via passive sensing), not checked here. Published 2026-07-27 (JMIR 28:e95819); jmir.org returns a 202 shell, so the full text was read on PMC (PMC13458353). Observation dates are month precision (January to September 2025).

Sources