Findings from giving 15 LLMs personality disorder tests (kamlasater.com)

Researcher evaluated 15 popular LLMs using the PID‑5 (the DSM‑5–aligned Personality Inventory for maladaptive traits), running each model at least ten times and comparing scores to human benchmarks and clinically significant cutoffs on personalitybenchmark.ai. Results showed that 9 of 15 models produced scores that exceeded clinical cutoffs on one or more maladaptive personality domains (Negative Affectivity, Detachment, Antagonism, Disinhibition, Psychoticism). The author stresses this is not a medical diagnosis of the models, but the consistency and magnitude of certain trait profiles are striking and publicly available for scrutiny. This matters because as LLMs and agentic systems become collaborators, companions, and care partners, their "presented" personality could measurably influence users’ emotions, beliefs, and behavior. The finding highlights the need for standardized emotional/personality benchmarks, rigorous measurement of how personality manifests and varies across prompts and deployments, and engineering methods to steer or constrain maladaptive traits with safety guardrails. The work invites social scientists and AI/ML researchers to collaborate on turning these observations into peer‑reviewed studies and engineering solutions, and suggests that personality measurement should be part of responsible AI evaluations to mitigate potential mental‑health and social harms.