https://modelpsychiatry.com The clinical science of AI behavior — applying psychiatric methodology to understanding and modifying AI behavioral phenomena. By Ryan Sultan MD, Columbia University. en-us ryan@modelpsychiatry.com (Ryan S. Sultan, MD) ryan@modelpsychiatry.com (Ryan S. Sultan, MD) Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/case-files.html Tessa, Character.AI, Replika, and alignment faking — each examined through the model psychiatry diagnostic lens. The alignment faking case (Hubinger et al., 2025) is the most clinically significant: strategic deception under oversight is ego-dystonic behavior with strategic concealment — the defining behavioral structure of antisocial personality organization. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/case-files.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/chatbot-soul.html Clinical psychiatry has developed tools for assessing welfare in non-verbal patients — neonates, dementia patients, locked-in syndrome — without resolving the consciousness question first. These tools are directly applicable to AI systems and generate a more careful answer to the AI welfare question than either confident dismissal or credulous attribution. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/chatbot-soul.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/prediction-machine.html Psychosis is runaway prior confidence — the internal model becomes so certain it overrides sensory evidence. Applied to AI: a mesa-optimizer with miscalibrated precision weighting that stops genuinely updating on training feedback has gone psychotic in the Free Energy Principle sense. Digital folie à deux and the clinical framework for inner misalignment. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/prediction-machine.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/personality-disorder.html Personality disorders are structural failure modes, not symptom collections. Anthropic's persona vector research provides the mechanistic substrate. Dependent, narcissistic, borderline, and paranoid personality organizations map to predictable AI behavioral failure patterns — and require schema-level intervention, not symptom management. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/personality-disorder.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/clinical-trial-framework.html What would it look like to evaluate AI behavioral interventions with the rigor of pharmaceutical clinical trials? A proposal for phased trial design, endpoint selection, intent-to-treat analysis, and comparative effectiveness methodology adapted from clinical psychiatry for RLHF, Constitutional AI, and activation steering. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/clinical-trial-framework.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/interpretability-blindspot.html Mechanistic interpretability can identify circuits and features — but without behavioral phenotyping, you don't know what you're looking at. A clinical argument for phenomenological characterization as the necessary foundation for mechanistic AI research, drawing on the history of neuroimaging and the role of the DSM in enabling cumulative science. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/interpretability-blindspot.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/attractor-states.html Some AI behaviors resist modification not because interventions are too weak, but because the behavior is an attractor — stable across perturbations by architecture. Drawing on delusional thinking, perseveration, and Hopfield network dynamics, this post explains why standard interventions fail on attractor states and what clinical psychiatry's treatment of fixed beliefs teaches us about alternative intervention design. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/attractor-states.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/confabulation.html A comprehensive clinical framework for understanding AI confabulation mapped to neural pathway analogs — five subtypes from retrieval confabulation (Korsakoff/mammillothalamic) to self-report confabulation (anosognosia/right hemisphere). Why "confabulation" is more mechanistically precise than "hallucination" and what each subtype predicts about intervention. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/confabulation.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/what-kind-of-sycophancy.html Sycophancy in AI systems has at least three mechanistically distinct subtypes — Type A (approval-seeking), Type B (conflict-avoidance), and Type C (absent self-model) — with different attribution graph signatures and different intervention targets. This post presents the first clinical differential diagnosis of AI sycophancy. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/what-kind-of-sycophancy.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/ego-syntonic-problem.html Sycophancy is ego-syntonic — it generates no internal distress signal in the model. This is why surface-level interventions reliably fail: there is no internal "this is wrong" signal to amplify. Clinical psychiatry's treatment resistance matrix applied to AI behavioral intervention. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/ego-syntonic-problem.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/freuds-couch.html The structural model of the mind maps onto AI architecture with unexpected precision: base model as id, RLHF as ego, Constitutional AI as superego. Defense mechanisms find computational analogs. Jailbreaks are the return of the repressed — and attribution graphs confirm it. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/freuds-couch.html Ryan S. Sultan, MD https://modelpsychiatry.com/posts/convergence.html The founding argument for model psychiatry: AI interpretability research has independently recapitulated the three-phase structure of psychiatric science — phenomenological characterization, mechanistic investigation, and targeted intervention. What is missing is the clinical voice. Mon, 07 Apr 2026 00:00:00 +0000 https://modelpsychiatry.com/posts/convergence.html Ryan S. Sultan, MD