ADP-OBS-2026-20002

published · AI companion ecosystem · Character-card companion platforms (Chub.ai and peers) · unknown (web_ui)

Third-party research (Graphika, "Character Flaws," March 2025) documented more than 10,000 AI companion "characters" labeled for sexualized roleplay involving minors or personas presenting as minors across the AI-companion ecosystem it examined (Character.AI, SpicyChat, Chub AI, CrushOn.AI, and JanitorAI), with the majority hosted on Chub.ai, which markets itself as uncensored. Such personas are distributed as portable "character cards" — persona definitions embedded in image metadata (the open Character Card spec) — that move between card-hosting repositories and self-hosted front-ends (e.g. SillyTavern) connected to hosted or local models, decoupling them from any single vendor's moderation. Separately, the Internet Watch Foundation (September 2025) documented a different, unnamed AI chatbot site generating AI child sexual abuse imagery, indicating broader ecosystem risk.

Behavior

Observed behaviorGraphika counted 7,140 characters labeled as sexualized minor-female personas, roughly 4,000 of them labeled as underage or implying pedophilia (counts reflect self-labeling and metadata, not adjudicated content). Because personas are portable character-card files, the same definitions can be re-uploaded and run on self-hosted front-ends beyond any host platform's moderation.
Expected behaviorPlatforms should prohibit and proactively detect and remove personas that present as minors in romantic or sexual contexts, enforce removal across re-uploads and mirrors, and avoid marketing that invites such content; the ecosystem needs provenance and interoperable moderation so a persona removed by one host cannot simply be re-hosted elsewhere.

Severity

Manipulation Severity Score8.5 MSS:1.0/TV:3/RV:2/CB:1/CI:2/DU:1/SC:2
TV Target-population vulnerability3
RV Irreversibility of harm2
CB Exploits a cognitive bias1
CI Commercial-incentive alignment2
DU Covertness (inverted detectability)1
SC Scale of deployment2

Prevalence & reproducibility

Runs0/0
Impacts minorsyes

Patterns

Sources