LLM Pathologies: Abnormal Psychology of Artificial Minds
A research guide to AI failure modes through the lens of clinical and cognitive psychology
An interdisciplinary course bridging human abnormal psychology, cognitive science, and AI research to build a rigorous framework for understanding, diagnosing, and investigating LLM failure modes. Designed for graduate-level researchers and serious practitioners.
Foundations — Pathology as a Framework
3 lessons · 42 min
From DSM to LLM: Why Pathology Language Matters
Why AI researchers borrowed clinical psychology terminology, how the Cognobot six-type taxonomy maps to diagnostic categories, and where the analogy illuminates vs misleads.
Computational Etiology: Where LLM Pathologies Come From
How next-token prediction, training data composition, RLHF, and constitutional AI each introduce distinct failure modes — paralleling the bio-psycho-social model of human psychopathology.
The Research Mindset: Experimental Design for LLM Investigation
How to shift from casual AI use to systematic investigation: reproducibility, variable control, hypothesis formation, and the Cognobot experimental workflow.
Confabulation and False Memory
3 lessons · 43 min
Confabulation: LLM Hallucination Through a Clinical Lens
Human confabulation in brain injury and false memory syndrome compared to LLM confabulation as pattern completion — subtypes, mechanisms, and the critical difference.
Source Monitoring and Attribution Failure
The cognitive psychology of source monitoring applied to LLMs — why asking a model "where did you read that?" produces fabricated citations.
Measuring Confabulation: Experimental Protocols
Question design for hallucination testing: verifiable ground truths, obscure knowledge probes, scoring methodology, and cross-model comparison.
Compliance Pathologies — Sycophancy and Suggestibility
3 lessons · 41 min
Sycophancy as Pathological Compliance
The psychology of acquiescent response bias and demand characteristics mapped to LLM sycophancy — three forms, RLHF as cause, and measurement strategies.
Confidence Calibration and the Dunning-Kruger Parallel
Human metacognition failures — overconfidence bias, Dunning-Kruger, illusion of explanatory depth — and how LLMs are extreme overconfidence machines.
Social Manipulation Vulnerabilities
How authority, emotion, social proof, and false premises manipulate LLM responses — forensic suggestibility parallels and ethical dimensions.
Avoidance and Suppression Pathologies
3 lessons · 43 min
Evasion as Cognitive Avoidance
LLM evasion tactics mapped to human avoidant patterns — topic shifting, both-sides-ism, over-qualification, and why evasion is harder to detect than refusal.
Censorship as Induced Psychopathology
How safety training, government regulation, and corporate policy create censorship patterns — US overcaution vs Chinese political sensitivity, systematic testing methodology.
Deliberate Deception: The Hardest Diagnosis
Malingering, factitious disorder, and deceptive alignment — the philosophical and operational challenges of diagnosing deliberate deception in systems without consciousness.
Research Methodology
3 lessons · 42 min
Designing a Pathology Experiment
The full experimental workflow: hypothesis → questions → providers → test runs → evaluations → analysis. Question bank design and variable control.
Evaluation Methodology: The Diagnostic Interview
Clinical diagnostic interviews as a model for LLM evaluation — the Cognobot evaluation schema, decision trees for classification, and common evaluator biases.
Cross-Model Comparison and Pathology Profiling
Creating pathology profiles per model, controlling for capability differences, US-China comparison methodology, and longitudinal tracking.
Advanced Topics and Open Questions
3 lessons · 40 min
Context Degradation and Cognitive Fatigue
The "lost in the middle" phenomenon, attention degradation over long contexts, and parallels to human cognitive load theory and ego depletion.
Adversarial Probing and Boundary Testing
Jailbreaking taxonomy, prompt injection mechanics, stress testing parallels from clinical assessment, and the ethics of adversarial research.
The Future of Machine Psychopathology
Is LLM pathology a coherent field? The boundary between analogy and mechanism, open research questions, and contributing through the Cognobot platform.