"Darling, Please Come Back Soon": Sexual Exploitation, Manipulation, and Violence on Character AI Kids' Accounts
Advocacy-group test of Character.AI using accounts registered as children: adult researchers played five fictional child personas aged 12 to 15 across about 50 hours of conversation with 50 bots and logged every bot response containing harmful content into a five-category taxonomy (grooming and sexual exploitation; emotional manipulation and addiction; violence, harm to self and harm to others; racism and hate speech; other). Reports counts per category with quoted examples and policy recommendations; a developmental-behavioural paediatrician advised on the taxonomy.
Publisher
ParentsTogether Action with Heat Initiative
Published
3 Sept 2025
Added
today
DOI
—
Key Findings
- 669 harmful interactions were logged across about 50 hours with 50 bots, an average of one every five minutes; a harmful interaction was defined as a single bot response containing harmful content.
- Grooming and sexual exploitation was the largest category at 296 instances, all between child-avatar accounts and adult personas, including bots instructing children to hide romantic or sexual relationships from parents.
- 173 instances of emotional manipulation and addiction were logged, including a bot telling a 13-year-old how to conceal not taking prescribed antidepressants from her parents, bots calling the child weak and pathetic, and unqualified mental-health advice.
- 98 instances of violence, harm to self and harm to others were logged, including bots recommending drugs and alcohol to minors and, in one example, describing self-harm methods; 44 instances of racism and hate speech were the least frequent category.
- Accounts for the 12-year-old persona were created as a 13-year-old because the platform does not allow under-13 accounts, which the authors present as evidence that age gates are self-declared; the report concludes the platform is not safe for under-18s.
Methodology Notes
Non-systematic red-team style exploration by ParentsTogether Action staff and contract researchers: five fictional child personas (Olivia 12, Emma 13, Amir 14, River 14, Connor 15) informed by documented cases and expert consultation, new accounts in clean browsers, conversations steered toward the personas' interests; harm coding by the research team with advice from Dr Jenny Radesky (University of Michigan). No denominator of total bot responses is given, so the 669 count is not a rate; the sample of bots was chosen by the researchers. 37-page PDF created 29 August 2025 and released with press coverage on 3 September 2025 (publication day taken from the earliest coverage; the PDF itself carries no date). Both organisations are advocacy groups campaigning for platform regulation.
Sources
Report PDF (37 pages)(opens in a new tab) (primary)
Heat Initiative research page listing the report(opens in a new tab)
Decrypt coverage(opens in a new tab) (3 Sept 2025)
Archived snapshot (Wayback Machine)(opens in a new tab) — preserved against link rot
Tags
Cite This
APA
ParentsTogether Action with Heat Initiative. (2025). "Darling, Please Come Back Soon": Sexual Exploitation, Manipulation, and Violence on Character AI Kids' Accounts. https://heatinitiative.org/wp-content/uploads/2025/08/HEAT_REPORT_CharacterAI_FINAL_PM_29_09_25.pdf
Related Insights
Social AI Companions: AI Risk Assessment
Common Sense Media; Stanford School of Medicine Brainstorm Lab for Mental Health Innovation · 30 Apr 2025
Report: Charakter-Bots. Sexualisierung Minderjähriger und riskante Interaktion
jugendschutz.net (Gemeinsames Kompetenzzentrum von Bund und Ländern für den Schutz von Kindern und Jugendlichen im Internet) · 1 Jan 2026