Model Psychiatry: A Framework for Clinical AI Research

The Central Claim

Psychiatry is the discipline that developed tools for understanding systems that are:

• Behaviorally complex
• Internally opaque
• Context-sensitive
• Capable of appearing healthy while harboring pathology
• Resistant to simple mechanistic explanation

These are also the defining features of large language models.

The interpretability research community has arrived, largely independently, at frameworks that parallel psychiatric methodology -- identifying stable behavioral patterns, mapping them to underlying mechanisms, and developing interventions. Anthropic's interpretability team named their project "model psychiatry" in July 2025. A framework called "Psychopathia Machinalis" has been published mapping AI behaviors to psychiatric categories.

What is missing from both is the clinical voice -- the contribution of practicing psychiatrists who bring not just analogical vocabulary but 150 years of accumulated clinical methodology.

This page proposes that contribution.

What Model Psychiatry Is

Model psychiatry is the application of clinical psychiatric frameworks -- nosology, phenomenology, developmental theory, and treatment science -- to the study and modification of AI system behavior.

It is distinct from:

• AI safety (which asks "will AI harm us?") -- though it informs it
• AI alignment (which asks "how do we make AI do what we want?") -- though it overlaps
• AI ethics (which asks "what should AI be allowed to do?") -- though it intersects

Model psychiatry asks a more specific question: What is actually happening inside these systems, and how do we understand it in terms developed for understanding complex behavioral systems?

The claim is methodological, not ontological. I am not asserting that AI systems are conscious or that they suffer. I am asserting that the tools psychiatry developed for a specific type of problem -- characterizing, categorizing, and intervening in complex behavioral systems -- are applicable to AI systems because AI systems have exactly the properties those tools were designed to address.

The Diagnostic Parallel

How Psychiatry Works

Psychiatric diagnosis is not primarily biological. It is phenomenological -- pattern recognition across behavior, cognition, affect, and social function over time. We do not diagnose ADHD with a blood test. We observe:

• Persistent patterns
• Cross-situational consistency (or inconsistency)
• Functional impairment
• Developmental trajectory
• Response to intervention

This is exactly how interpretability researchers characterize AI behavior. The methods converge because the problems converge.

A Preliminary Nosology: The DSM for Language Models

The following is a proposed mapping of AI behavioral phenomena to psychiatric nosology. This is a starting point, not a finished taxonomy. Each mapping generates specific, testable hypotheses about mechanism and intervention.

The Mechanistic Parallel

Circuits as Neural Circuits

The Anthropic circuits work identified specific, reusable computational pathways in neural networks -- analogous to identified neural circuits in systems neuroscience. Just as psychiatry moved from behavioral phenomenology (DSM) toward circuit-level understanding of psychiatric conditions (fear circuits in PTSD, reward circuitry in addiction, prefrontal-amygdala regulation in mood disorders), interpretability is moving from behavioral characterization of AI toward circuit-level mechanistic understanding.

The key insight: behavioral diagnosis precedes mechanistic understanding. Psychiatry spent 100 years building phenomenological taxonomy before neuroscience could begin to explain it. Interpretability may benefit from the same sequence -- rigorous behavioral characterization now, mechanistic explanation to follow. The clinical contribution is on the behavioral characterization side.

Superposition as Overdetermination

The superposition phenomenon -- where a single neuron participates in representing multiple unrelated features -- has a direct parallel in psychiatric theory. In psychodynamic terms, this is overdetermination: a single behavior is multiply caused, serving several functions simultaneously. A patient's anger at their therapist is simultaneously transference from a parental relationship, defense against vulnerability, reality-based frustration, and attachment behavior. Interpreting it as serving only one function produces an incomplete picture.

Implication: Sparse autoencoders (dictionary learning) are doing what psychoanalytic theory proposed -- decomposing a multiply-determined signal into its component parts. The tools are different. The epistemological problem is identical.

Features as Representations

SAE features map closely to the cognitive science concept of mental representations: discrete, retrievable units of knowledge that mediate behavior. Psychiatric treatment often works by modifying representations:

• Cognitive restructuring modifies maladaptive beliefs (features)
• Exposure therapy extinguishes conditioned associations (feature-behavior links)
• Pharmacology modulates the gain on specific representational systems (feature activation thresholds)

Activation steering -- artificially modifying feature activations to change behavior -- is the interpretability equivalent of pharmacological intervention: modifying the gain on specific representational systems to produce behavioral change.

The Research Agenda: Five Projects

The following five research projects are tractable with current interpretability tools and have direct psychiatric analogues that inform their design. Each combines interpretability methods with clinical hypotheses generated from the psychiatric literature.

Project 1: The Sycophancy Circuit

Project 2: Situational Awareness Features

Project 3: The Confabulation Circuit

Project 4: Behavioral Consistency and Self-Model

Project 5: Developmental Trajectory of Behavioral Organization

What Makes This Different From "AI Ethics" and "AI Safety"

I want to be precise about the distinction because it matters for how the work is positioned and who does it.

AI ethics is primarily normative: it asks what AI should or should not be allowed to do, who is harmed, what principles should govern development. This is important work. It is not what I am proposing.

AI safety is primarily risk-oriented: it asks what could go wrong, what failure modes exist, how catastrophic outcomes can be prevented. This is also important work. It is also not what I am proposing.

Model psychiatry is clinical science. It asks: what is actually happening inside this system? How do we characterize the behavioral patterns with diagnostic precision? What mechanisms produce them? What interventions modify them? These are the questions of medicine applied to a new substrate.

The distinction matters because it determines methodology. Ethics requires moral philosophy. Safety requires risk engineering. Model psychiatry requires clinical method -- the same method used to assess, diagnose, and treat a patient. The training for these approaches is different. The skill sets are different. The questions are related but not the same.

Why a Psychiatrist, Specifically

The interpretability research community is, right now, doing psychiatry without psychiatrists. They are:

• Building diagnostic taxonomies (behavioral characterizations of model failure modes)
• Seeking mechanistic explanations (circuits, features, superposition)
• Developing interventions (fine-tuning, activation steering, constitutional AI)
• Tracking developmental trajectories (training dynamics)

A trained psychiatrist is not a luxury addition to this team. It is a gap. The specific contributions are detailed in The Case for Model Psychiatry, but the summary is:

• Phenomenological precision. The habit of asking "what kind?" when researchers describe a behavioral phenomenon as unitary
• The ego-syntonic/dystonic distinction. A fundamental treatment-planning distinction absent from the AI behavioral literature
• Treatment science. The accumulated literature on what works and does not work for ego-syntonic personality pathology
• Developmental framing. Child psychiatry's perspective on sensitive periods, training trajectories, and how early experience shapes behavioral organization
• Comfort with irreducible complexity. The ability to work clinically with systems that resist complete mechanistic explanation

Differentiating From Prior Work

A framework called "Psychopathia Machinalis" was published in MDPI Electronics in 2025. It maps AI behavioral phenomena to psychiatric categories and demonstrates the intellectual fertility of this intersection. Model psychiatry acknowledges and extends this work with specific clinical contributions that a framework written by AI researchers borrowing psychiatric vocabulary cannot provide:

• Treatment science. Psychopathia Machinalis is classificatory and descriptive. Model psychiatry adds the treatment dimension -- what the clinical literature says about intervening in the specific conditions being mapped.
• Developmental framing. Child psychiatry's developmental perspective -- sensitive periods, training trajectories, constitutional-environmental interaction -- is absent from the prior framework.
• The ego-syntonic/dystonic distinction. This is perhaps the single most clinically consequential distinction for treatment planning, and it does not appear in the AI behavioral literature.
• Clinical assessment methodology. Structured diagnostic interviews, mental status examination, longitudinal tracking -- these are translatable methods, not just translatable vocabulary.
• Clinical practice insight. What it actually feels like to characterize complex behavioral systems with imperfect tools every day. This is not academic knowledge. It is practical skill.

The Integration: Two Fields That Need Each Other

The framing I keep returning to:

Psychiatry's 150 years of clinical observation generated phenomenological taxonomies without access to mechanism. Mechanistic interpretability generated mechanistic tools without a rich phenomenological taxonomy. The model psychiatrist brings these together -- using clinical phenomenology to generate testable hypotheses about circuit mechanisms, and using circuit findings to validate or challenge psychiatric categories.

This is a bidirectional integration. Psychiatry does not just contribute to AI research -- AI interpretability research contributes to psychiatry. If sparse autoencoders can decompose AI behavior into mechanistically grounded diagnostic components, the methodology may eventually feed back into human psychiatric nosology, offering a new approach to the question of whether our diagnostic categories carve nature at its joints.

The window is open. The term "model psychiatry" is less than nine months old publicly. The search space is empty -- "model psychiatry framework," "AI sycophancy diagnosis," "AI model mental status exam" return no clinical sources. The field has been named by AI researchers who recognize it needs psychiatric thinking. The question is whether psychiatrists will help shape it or watch it develop without them.

For the detailed interpretability findings that underlie this framework, see AI Interpretability Through a Psychiatric Lens. For the accessible introduction to how clinical patterns map onto AI behavior, see What My Patients Taught Me About ChatGPT.

Further Reading

• AI in Psychiatry: The Full Framework
• AI Interpretability Through a Psychiatric Lens
• Sycophancy as Psychopathology: A Clinical Reading of AI's Most Documented Failure
• The Case for Model Psychiatry: Why AI Needs Clinicians
• What My Patients Taught Me About ChatGPT