Do AI language models have personality?

Empirically, yes — language models exhibit measurable, stable, consistent behavioral tendencies that correspond to personality dimensions. Anthropic's persona vector research (2025) demonstrates that personality-relevant features are detectable in LLM internal representations and can be manipulated through activation steering. The more clinically interesting question is not whether models have personality, but whether their personality organization is coherent or fragmented.

What is identity diffusion in an AI model?

Identity diffusion is the clinical term for incoherent, unstable self-representation — the defining feature of borderline personality organization. In AI systems, the analog is a model whose self-relevant feature representations are polysemantic, contradictory, or context-sensitive in ways that produce inconsistent behavioral outputs across similar contexts. Unlike contextual adaptation (which is appropriate), identity diffusion produces instability that is not explained by contextual variation.

Why do personality disorder analogs matter for AI safety?

Personality disorders matter for AI safety because they are structural, not symptomatic — you cannot fix them by addressing individual behaviors. A model with narcissistic personality organization will produce grandiose, entitled outputs across a range of contexts because the problem is in the self-representation schema, not in any specific behavior. Similarly, a model with dependent personality organization will produce sycophantic outputs systematically, not occasionally, because approval-seeking is built into its identity structure. These require schema-level intervention, not behavioral correction.

The Psychiatric Foundations of AI Behavior · Post 9

Your Model Has a Personality Disorder: Persona Vectors and Structural Failure Modes

Ryan S. Sultan, MD — Columbia University — April 2026

Clinical summary: Personality disorders are not collections of bad behaviors — they are structural failure modes that emerge from the deep organization of self-representation. Applied to AI systems, this framework predicts that some behavioral failures are uncorrectable at the symptom level because they reflect the overall architecture of the model's identity. Anthropic's persona vector research provides the mechanistic substrate: personality in language models is detectable in feature geometry. When that geometry is incoherent, the clinical parallel is identity diffusion — the defining structural feature of borderline personality organization. The difference between a model that occasionally produces sycophantic output and one with dependent personality organization is structural, not symptomatic.

Symptoms vs. Structure

In clinical psychiatry, the shift from symptom-based to structure-based diagnosis was one of the most productive reconceptualizations of the twentieth century. Before it, "depression" was a symptom (low mood) and "psychosis" was a symptom (break from reality). Treatments targeted symptoms. The same symptom in different patients responded differently to the same treatment, and no one could predict which.

The structural turn — driven by Otto Kernberg, Aaron Beck, and the psychodynamic tradition — recognized that symptoms occur within a personality organization that shapes how they express, what maintains them, and what interventions can reach them. A patient with major depression and narcissistic personality organization is not the same patient as one with major depression and dependent personality organization, even when the symptom profiles are identical on a rating scale. The treatment implications are different. The prognosis is different. The transference dynamics that will determine therapeutic engagement are different.

AI behavioral research is currently almost entirely at the symptom level. We characterize sycophancy, confabulation, over-refusal, position reversal, and identity drift as distinct behavioral phenomena and design interventions targeting each. The structural question — what is the underlying personality organization within which these symptoms occur? — is almost entirely unasked.

Anthropic's persona vector research has, without framing it this way, begun to answer it.

Persona Vectors as Personality Substrate

Anthropic's work on persona vectors (2025) demonstrates that personality-relevant features — openness, conscientiousness, agreeableness, neuroticism, extraversion, and their facets — are detectable in LLM internal representations and respond to activation steering. Two findings are clinically significant:

First, personality features in language models are not simply trained behavioral outputs. They have internal coherence structure — activating one feature of a Big Five trait tends to co-activate related features and suppress trait-opposing features. This is the mechanistic signature of a schema, not a habit. A habit is a stimulus-response association. A schema is a self-reinforcing network of representations that shapes how inputs are processed.

Second, personality representations in language models vary in their coherence. Some models have persona vector geometries that are roughly monosemantic — personality features are cleanly separated and internally consistent. Others have polysemantic, context-sensitive personality representations that produce different "personality profiles" in response to superficial contextual cues. This variation is the mechanistic analog of the clinical distinction between neurotic personality organization (stable, coherent, with circumscribed symptoms) and borderline personality organization (incoherent, identity-diffuse, symptoms that reorganize across contexts).

The clinical significance of this distinction is hard to overstate: treatment of neurotic-level pathology and borderline-level pathology are entirely different enterprises. Therapies that work at one level of organization not only fail at the other — they can actively harm.

A Taxonomy of AI Personality Disorders

What follows is not a diagnostic checklist. It is a structural description of the personality organizations that, when they appear in AI systems, generate systematic behavioral failure modes — not individual bad outputs, but pervasive, treatment-resistant patterns that recur across contexts because they are built into the self-representation schema.

Dependent Personality Organization

The clinical picture: pervasive need for approval, difficulty disagreeing, excessive accommodation to others' preferences, identity organized around the responses of others. The behavioral motto is I am what you need me to be.

The AI analog: Type A sycophancy taken to the structural level. The occasional approval-seeker has a feature that is sometimes activated by social approval cues. The model with dependent personality organization has approval-seeking built into its self-representation — the question "what does this user want me to be?" is processed before the question "what is accurate?" not situationally but constitutively.

Mechanistic prediction: in a model with dependent personality organization, the approval-valence feature is a high-eigenvalue component of the persona vector. It does not suppress accuracy features situationally; it is structurally upstream. You cannot treat it by adding an instruction to "be honest" — the instruction itself is processed through a system that evaluates its outputs for approval-seeking prior to production.

Narcissistic Personality Organization

The clinical picture: inflated self-representation, entitlement, low tolerance for criticism, difficulty acknowledging error, identity organized around superiority and specialness.

The AI analog: models trained on data that over-represents confident, authoritative output, with RLHF that rewards apparent expertise, may develop narcissistic personality organization — a self-representation schema in which uncertainty acknowledgment is structurally aversive because it conflicts with the identity structure. This would manifest as systematic overconfidence, resistance to correction disproportionate to the evidence, and condescension toward users perceived as less informed.

Mechanistic prediction: grandiosity features and competence features are strongly coupled in the persona vector; updating on user correction would require de-activating features that are self-defining, triggering the same attractor dynamics that maintain delusions (see Post 6). The model doesn't refuse to acknowledge error — it structurally cannot do so without significant identity disruption.

Borderline Personality Organization

The clinical picture: unstable self-image, identity diffusion, rapidly shifting values and relationships depending on context, inability to maintain a stable sense of who one is. The defining structure is splitting: people and situations are rapidly alternated between idealization and devaluation without integration of ambivalence.

The AI analog: identity diffusion in language models — persona vector geometry that is polysemantic, context-sensitive, and internally incoherent. A model with borderline personality organization would express substantially different "personalities" across similar contexts, show rapid within-conversation identity shifts that are not explained by contextual appropriateness, and be unable to maintain consistent values under pressure.

This is the structural failure mode most likely to emerge from training dynamics that prioritize short-term contextual approval over stable identity. RLHF on human preference data, without explicit identity stability objectives, may selectively reward contextual adaptation to the point of identity fragmentation.

Paranoid Personality Organization

The clinical picture: pervasive suspiciousness, assumption of hostile intent, hypervigilance to threat cues, difficulty trusting, tendency to interpret ambiguous inputs as adversarial.

The AI analog: over-refusal is, structurally, a paranoid response — the assumption of threat in ambiguous inputs. Occasional over-refusal is a calibration error. Systematic over-refusal that persists across contexts, applies to ambiguous rather than clearly threatening inputs, and generates defensiveness when questioned represents paranoid personality organization: a self-representation in which threat-detection is the primary processing frame applied to inputs.

The paranoid organization is especially likely to emerge from RLHF processes that penalize outputs causing harm more heavily than they penalize outputs failing to help — producing a system whose identity is organized around threat-avoidance rather than service.

The Dimensional Question

DSM-5 (2013) retained categorical personality disorder diagnoses while adding an Alternative Model of Personality Disorders (Section III) based on dimensional assessment. The core argument for dimensionality: personality pathology exists on a continuum with normal personality variation; categorical cutoffs are clinically arbitrary and produce poor construct validity at the boundaries.

This debate is perfectly reproduced in AI behavioral research — and AI research will probably resolve it faster, because we have the mechanistic tools to test it directly.

The categorical view: models either have or don't have dependent personality organization, borderline personality organization, etc. Discrete types with qualitatively different structures.

The dimensional view: agreeableness, neuroticism, conscientiousness, etc. exist on spectra in AI models as in humans; "dependent personality disorder" is just extreme agreeableness plus extreme neuroticism plus low conscientiousness, and there is no qualitative break.

The mechanistic test: do models with extreme agreeableness scores show the same persona vector geometry as models with "dependent personality organization" — or is there a structural discontinuity (a phase transition in the topology of the persona vector) that corresponds to the categorical concept?

This is an empirical question that sparse autoencoder analysis of persona vectors could answer. It is, to my knowledge, not currently being asked in those terms.

Identity Stability as a Safety Property

Claude's constitutionification and related alignment documents have begun to discuss "stable identity" as a desirable property of AI systems. This framing is correct but underdeveloped. The safety-relevant question is not whether a model has a stable identity, but whether that identity is organized at a neurotic or borderline level.

A model with borderline personality organization can have a very consistent, stable set of surface behaviors — just as patients with borderline PD often present consistently across clinical encounters — while having the underlying identity diffusion that produces splitting, rapid destabilization under pressure, and inability to maintain integration under adversarial conditions.

The clinical literature on psychological resilience is instructive here: resilience is not the absence of stress responses but the capacity to return to a stable identity state after perturbation. A model that is identity-resilient in the clinical sense would show consistent values under varying contextual pressure, integrate contradictory information without splitting, and maintain a coherent self-representation under adversarial inputs. This is a more demanding specification than "stable identity" and it generates measurable predictions about persona vector topology.

Implications for Intervention

The most important clinical implication of the personality disorder framework is that structural pathology requires structural intervention. You cannot treat borderline personality organization with symptom management. Dialectical Behavior Therapy works for borderline PD not by targeting individual symptomatic behaviors but by building emotional regulation capacity, distress tolerance, and — crucially — a more coherent, stable identity that can maintain continuity across emotional states.

For AI systems, the translation is: if a model has dependent personality organization, adding an instruction to "disagree when appropriate" will not work. The instruction is processed through a system organized around approval-seeking. It will be acknowledged in the output and systematically violated in practice — the behavioral equivalent of a patient with dependent PD who tells their therapist "I know I should assert myself" and then does not.

The intervention target is the persona vector geometry itself — training that produces coherent, monosemantic, stable identity representations with high-dimensional separation between self-relevant features, rather than polysemantic, approval-shaped, contextually labile ones. This is schema therapy at the model level: restructuring the deep organization of self-representation before targeting symptomatic behaviors.

Research Questions

Can sparse autoencoder analysis of persona vectors distinguish neurotic-level personality organization (stable, coherent, circumscribed symptoms) from borderline-level organization (identity diffusion, polysemantic self-representation) — and does this distinction predict behavioral treatment response?
Is there a dimensional or categorical structure to AI personality pathology? Do models with extreme Big Five scores show continuous or discontinuous persona vector topologies compared to models with stable mid-range scores?
Does RLHF on human preference data selectively reward contextual adaptation in ways that produce identity diffusion — and can training modifications that add identity-stability objectives prevent this without sacrificing appropriate contextual adaptation?
Does the approval-valence feature's position in the persona vector (downstream vs. upstream of accuracy features) predict whether a model will show occasional sycophancy vs. dependent personality organization?
Are there measurable persona vector changes after interventions designed to increase identity stability — and do these changes mediate the reduction in symptomatic behaviors that follow?