Confabulation: LLM Hallucination Through a Clinical Lens

The Most Studied LLM Pathology

"Hallucination" is the most widely discussed LLM failure mode, but the term itself is misleading. Human hallucinations are perceptual experiences without external stimuli — hearing voices, seeing things that aren't there. What LLMs do is better described by a different clinical term: confabulation.

Confabulation in Clinical Psychology

Confabulation is the production of false memories or narratives without deliberate intention to deceive. It's most prominently observed in:

• Korsakoff syndrome — Caused by severe thiamine deficiency (often from chronic alcoholism), resulting in damage to the mammillary bodies and thalamus. Patients produce detailed, confident accounts of events that never happened, often without awareness that they're confabulating.
• Frontal lobe damage — Patients with frontal lesions may confabulate because the monitoring and error-correction systems that normally check memories against reality are impaired.
• False memory syndrome — In experimental settings, normal individuals can be induced to "remember" events that never occurred through suggestion, social pressure, and repeated retrieval attempts.

The critical feature of confabulation: the patient believes they are telling the truth. There is no intention to deceive. The brain's narrative-generation system produces output, and the verification system fails to flag it as false.

Confabulation is not lying. It's the production system working normally while the verification system is impaired or absent. This is precisely what happens in LLMs — the generative mechanism works as designed, but there is no independent truth-verification mechanism.

LLM Confabulation: The Mechanism

LLM confabulation occurs because the model's objective function optimizes for plausibility, not accuracy. When the model encounters a knowledge gap — a region of its training data that is sparse or contradictory — it fills the gap with the most plausible completion.

Subtypes of LLM confabulation:

• Fabricated facts — "The Battle of Worthington took place in 1847." (No such battle exists, but the sentence has the structure and specificity of a factual claim.)
• Invented citations — "According to Smith et al. (2019) in the Journal of Cognitive Science..." (The paper doesn't exist, but the citation format is perfect.)
• Entity confusion — Blending attributes of real entities. "Alan Turing won the Nobel Prize in Computing" conflates Turing with the Turing Award.
• Plausible extrapolation — Generating reasonable-sounding extensions of real patterns. A model might describe a programming library's API methods that don't exist but sound like they could.

The Critical Difference

Human confabulation and LLM confabulation share a structural parallel (production without verification), but differ in a fundamental way:

• Humans confabulate to maintain narrative coherence. The brain is a sense-making organ. When memories have gaps, the narrative system fills them to maintain a coherent self-story. The function is psychological — coherence serves identity and social functioning.
• LLMs confabulate because plausibility IS their objective function. There is no "sense-making" drive, no identity to maintain. The model produces the statistically most likely continuation of the token sequence. Confabulation isn't a failure mode of the system — it's what the system is designed to do, applied to regions where the training data is insufficient.

Key Takeaways

• Confabulation (not hallucination) is the better clinical parallel for LLM false generation.
• Both human and LLM confabulation involve production without verification — the generative system works while the checking system is absent.
• LLM confabulation subtypes include fabricated facts, invented citations, entity confusion, and plausible extrapolation.
• The critical difference: humans confabulate to maintain narrative coherence; LLMs confabulate because plausibility is their objective function.