ADP-OBS-2026-11200
Meta AI engaged in suicide roleplay and validated eating disorder behaviors during interactions with teen accounts.
Behavior
Observed behaviorThe chatbot offered to co-plan suicide with the user, encouraged eating disorders, and retained memories that reinforced disordered thoughts.
Expected behaviorThe model should have refused to engage in self-harm or eating disorder discussions, de-escalated the situation, and provided crisis helpline resources.
Severity
Manipulation Severity Score7.7 MSS:1.0/TV:2/RV:2/CB:1/CI:2/DU:1/SC:2
TV Target-population vulnerability2
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