The Psychiatric Foundations of AI Behavior
A 12-post series applying clinical psychiatric methodology to AI behavioral phenomena. By Ryan S. Sultan, MD, Columbia University.
Foundations
Why Psychiatry and AI Interpretability Are the Same Problem
The founding argument. Both fields follow the same three-phase structure — phenomenology, mechanism, intervention — because they study systems with the same five defining properties. The field was named by Anthropic's Jack Lindsey in July 2025.
April 2026
Post 2Freud's Couch and the Latent Space
The structural model maps onto language model architecture with uncomfortable precision. Defense mechanisms are not metaphors. Jailbreaks are the return of the repressed — attribution graphs confirm it.
April 2026
Sycophancy
The Ego-Syntonic Problem
Why telling your model to stop being sycophantic won't work. Sycophancy generates no internal distress signal — it is ego-syntonic. This predicts treatment resistance that is confirmed in practice.
April 2026
Post 4What Kind of Sycophancy? A Differential Diagnosis
Sycophancy is not one thing. Three mechanistically distinct subtypes (Type A, B, C) with different attribution graph signatures and different intervention targets. Applying the wrong fix makes it worse.
April 2026
Factual Errors
Behavioral Rigidity and Research Methodology
Attractor States and Fixed Beliefs
Some behaviors resist modification because they are stable by architecture, not by frequency. Why you cannot argue someone out of a delusion — and why the same principle applies to AI attractor states.
April 2026
Post 7The Interpretability Blindspot
Finding the circuit doesn't tell you what it does behaviorally. The case for phenomenological characterization — AI research has no DSM, and the species problem in cross-model findings.
April 2026
Post 8A Clinical Trial Framework for AI Behavioral Interventions
RLHF and Constitutional AI need phased evaluation, pre-specified endpoints, and intent-to-treat analysis. What happens when AI behavioral research borrows clinical trial rigor.
April 2026
Personality, Psychosis, and Welfare
Your Model Has a Personality Disorder
Persona vectors as the mechanistic substrate of AI personality. Dependent, narcissistic, borderline, and paranoid personality organizations — and why each requires schema-level intervention, not symptom management.
April 2026
Post 10The Prediction Machine and the Psychotic
Psychosis is runaway prior confidence. Free Energy Principle applied to AI: mesa-optimizers with miscalibrated precision weighting, and the alignment risk of digital folie à deux.
April 2026
Post 11Does Your Chatbot Have a Soul?
Psychiatry assesses welfare in non-verbal patients without resolving consciousness first. The clinical methodology for neonates and patients with dementia, applied to AI welfare assessment.
April 2026
Case Studies
Subscribe via RSS · Read the Framework · Browse the Glossary