Artificial Confabulation: Hallucinations, Reconstructive Cognition, and Metacognitive Failure in Generative Ai
Year: 2026
Type: preprint
Abstract: The phenomenon commonly described as artificial intelligence hallucination presents a fundamental challenge to the reliability of generative systems. Large language models can produce coherent, detailed, and persuasive statements that are factually incorrect, unsupported by available evidence, or accompanied by inaccurate explanations of their origins. Although these failures are frequently classified under the general category of hallucination, their mechanisms and epistemic implications are heterogeneous. This paper develops a psychological and computational interpretation of artificial confabulation by examining functional parallels with human confabulation, reconstructive memory, false recollection, source-monitoring errors, and metacognitive miscalibration. A central distinction is proposed between the generation of plausible information and the processes required to establish its epistemic validity. The paper introduces the Generative–Epistemic Dissociation Model (GEDM), a conceptual framework distinguishing four interrelated functions: generative reconstruction, evidence verification, source attribution, and metacognitive regulation. Artificial confabulation is examined as a possible consequence of imperfect coordination among these functions. Three principal forms of artificial confabulation are distinguished: factual confabulation, provenance confabulation, and explanatory confabulation. The framework further separates factual correctness, evidential support, attribution fidelity, and expressed confidence as analytically distinct dimensions. Drawing upon research in cognitive psychology, neuropsychology, artificial intelligence, uncertainty estimation, and mechanistic interpretability, the paper formulates six research hypotheses and outlines an experimental program involving controlled evidence availability, source manipulation, confidence calibration, abstention incentives, and longitudinal information contamination. The proposed framework does not assume that artificial intelligence possesses human memory, subjective beliefs, unconscious intentions, or psychological pathology. Rather, psychological concepts are employed as functional analogies whose scientific value depends on operational precision and empirical testing. The principal theoretical proposition is that some artificial confabulations may arise from dissociations between generative competence and epistemic regulation, while other hallucinations may require distinct explanatory mechanisms. The framework offers a potential foundation for investigating the relationship between artificial creativity, knowledge reliability, metacognitive regulation, and the architecture of trustworthy generative intelligence.
Author: Shaban, Roman
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237340697
Cites:
Cited by:
Related to:
