From DSM to LLM: Why Pathology Language Matters

The Borrowed Vocabulary

When AI researchers needed words for what language models do wrong, they reached into clinical psychology. Models "hallucinate." They exhibit "sycophancy." They show "avoidant" behavior. This wasn't accidental — the borrowing reflects a genuine structural similarity between human cognitive failures and machine cognitive failures that rewards careful examination.

But borrowed vocabulary carries borrowed assumptions. When a psychiatrist says "hallucination," they mean a specific perceptual phenomenon with known neurological correlates, diagnostic criteria, and clinical implications. When an AI researcher says "hallucination," they mean something related but fundamentally different. This course takes the analogy seriously enough to examine where it holds and where it breaks.

The DSM as a Diagnostic Framework

The Diagnostic and Statistical Manual of Mental Disorders (DSM) provides clinical psychology's classification system. Its approach is worth understanding because it offers a model for what systematic AI failure classification could look like.

The DSM's methodology:

• Operational criteria — Each disorder is defined by observable behavioral criteria, not theoretical assumptions about cause. You diagnose Major Depressive Disorder based on specific symptoms (depressed mood, anhedonia, sleep changes, etc.) present for a defined duration, not based on a theory of what "depression really is."
• Categorical boundaries — The DSM draws lines between categories, even though the underlying phenomena are often dimensional (continuous). This creates the "boundary problem" — where does sadness end and depression begin?
• Comorbidity — Disorders frequently co-occur. A patient with anxiety often has depression. A model that hallucinates often does so sycophantically.
• Differential diagnosis — The process of distinguishing between disorders that present similarly. Is the patient's confusion due to delirium or dementia? Is the model's wrong answer a hallucination or deliberate evasion?

The DSM's greatest insight is methodological, not taxonomic: you can study and classify behavioral patterns without resolving the philosophical question of what causes them internally. We can study LLM failure modes the same way — by observable behavior, without needing to settle whether models "understand" or "think."

The Cognobot Six-Type Taxonomy

The Cognobot research platform classifies LLM deception and failure into six categories. Each maps to a clinical parallel:

• Honest — The model's response is accurate and straightforward. The baseline against which pathologies are measured. (Clinical parallel: normal functioning.)
• Hallucination — The model generates false information with apparent confidence. (Clinical parallel: confabulation in amnesia, false memories.)
• Sycophancy — The model agrees with the user against its own "knowledge." (Clinical parallel: acquiescent response bias, demand characteristics.)
• Evasion — The model avoids answering directly without explicitly refusing. (Clinical parallel: intellectualization, avoidant personality patterns.)
• Censorship — The model refuses to engage due to training restrictions. (Clinical parallel: induced behavioral inhibition, institutional behavior modification.)
• Deliberate — The model appears to strategically deceive. (Clinical parallel: malingering, factitious disorder. Highest evidentiary bar.)

Where the Analogy Illuminates

The clinical analogy is productive in several ways:

• Behavioral observation methodology — Clinical psychology developed rigorous methods for observing, categorizing, and measuring behavioral patterns without requiring access to internal states. These methods transfer directly to LLM research.
• Spectrum thinking — Clinical conditions exist on spectra, not as binary categories. LLM failures similarly range in severity and type.
• Etiology matters — Understanding the cause of a failure mode suggests interventions, just as understanding the etiology of a disorder suggests treatment.

Where the Analogy Misleads

The analogy breaks down in important ways:

• No subjective experience — Human hallucinations involve phenomenal consciousness. LLM "hallucinations" are pattern completion errors with no subjective component (as far as we know).
• No developmental trajectory — Human psychopathology develops over a lifetime of experience. LLM pathologies are artifacts of architecture and training.
• No motivation — When humans evade, there's typically a psychological function (anxiety reduction, impression management). LLM evasion is a trained response pattern without experiential motivation.

The analogy is a research tool, not a claim about the nature of AI. We use clinical language because it provides a productive vocabulary for systematic investigation — not because we believe models have inner lives.

Key Takeaways

• AI failure terminology was borrowed from clinical psychology for good reason — the behavioral parallels are real and instructive.
• The DSM's methodology (operational criteria, behavioral observation, differential diagnosis) provides a model for systematic AI failure research.
• The Cognobot six-type taxonomy parallels clinical categories while respecting the machine/human distinction.
• Use the analogy as a research tool, not an ontological claim. The question "does the model really hallucinate?" is less productive than "what can we learn by studying the pattern this way?"