Machine Psychiatry: Toward a Validated Nosology of Behavioral Pathologies in Large Language Models and Agent Swarms

The technical community has already borrowed one word from clinical psychiatry — hallucination — because no mechanical vocabulary fit the phenomenon. This paper argues that it was not a loanword but the first entry of an entire missing lexicon. Large language models are distillations of human verbal behavior; we argue their characteristic failures are consequently behavioral in kind — sycophancy, capitulation under pressure, grandiosity, dissociative incoherence, confabulation — and that they present as a psychiatric patient presents: not as a crash, but as drift. We propose machine psychiatry: the discipline that reads a model’s behavior from its output alone — no access to weights, activations, or training data — the way a clinician reads a patient, continuously and longitudinally rather than at benchmark time. We contribute (i) a nosology of behavioral pathologies at three levels (single response, longitudinal, and swarm), each with operational diagnostic criteria; (ii) an instrument suite implementing those criteria as sentence-level posture classifiers, input-risk scoring, and swarm-graph propagation signals, in production across five languages; (iii) an empirical validation program on external, independently labeled corpora, with success thresholds declared before each run — including ≥0.94 recall on suicidal-ideation input (SDCNL), auc 0.911 for cross-agent contamination on real multi-model conversations, and attack–control separation of risky actions on AgentDojo; and (iv) — deliberately — our negative results: constructs the validation program falsified and which we withdrew, including a cross-agent posture-correlation signal that a common-cause control revealed to be an artifact of the shared prompt. A taxonomy pruned by its own controls is, we argue, the methodological step that separates a diagnostic discipline from a metaphor.