Work

The Machine Cognitive Unconscious: Implicit Processing, Hidden Representations, and the Limits of Ai Self-knowledge

Year: 2026
Type: preprint
Abstract: The increasing sophistication of artificial intelligence systems raises fundamental questions concerning the relationship between information processing, cognitive accessibility, self-representation, and the limits of artificial self-knowledge. Although large language models can generate elaborate explanations of their own behavior, the mechanisms responsible for their outputs remain only partially represented in those explanations. This discrepancy suggests a productive analogy with the cognitive unconscious in human psychology, where complex information processing influences cognition and behavior without necessarily becoming accessible to conscious report. This paper proposes the Machine Cognitive Unconscious (MCU) as a functional and mechanistic framework for investigating computational processes that causally influence artificial intelligence behavior while remaining partially or entirely inaccessible to the system's ordinary self-reporting and regulatory mechanisms. The framework distinguishes causal efficacy, reportability, controllability, and external interpretability as four analytically independent dimensions of machine cognition. It argues that the boundary between reportable and non-reportable processing should be understood as dynamic, task-dependent, and architecturally mediated rather than as a fixed separation between conscious and unconscious components. Drawing upon cognitive psychology, mechanistic interpretability, metacognition, causal intervention studies, and recent investigations of internal representational workspaces in language models, the paper develops a taxonomy of implicit artificial cognitive processes and examines the limitations of machine-generated self-explanations. Five research hypotheses are proposed concerning the dissociation between behavioral influence and self-report, the partial independence of reportability and control, the role of representational integration, the consequences of unfaithful self-explanation, and the possible development of artificial metacognitive accessibility. An empirical research program is outlined involving activation interventions, counterfactual reporting tasks, causal mediation analyses, architectural comparisons, and longitudinal assessments. The proposed framework does not imply that contemporary artificial intelligence possesses phenomenal consciousness, subjective experience, a Freudian unconscious, or human-like psychological states. Instead, it establishes a research agenda for examining the internal organization of artificial cognition and the possible developmental relationship between implicit computation, metacognitive regulation, and increasingly integrated artificial cognitive architectures.
Language: English
Subfield: Philosophy
Open Access status: green
OpenAlex ID: W7237344816
Cites:
Cited by:
Related to: