Work

From Black-Box Computation to Causal Accessibility: Toward a Theory of Emergent Cognitive Capabilities in Artificial Intelligence

Year: 2026
Type: preprint
Abstract: The increasing capabilities of artificial intelligence systems have intensified a fundamental question concerning the nature of machine intelligence: how can computational systems acquire cognitive capabilities that were neither explicitly programmed nor fully anticipated by their designers? Conventional explanations emphasize architectural complexity, parameter scaling, optimization, distributed representations, and the acquisition of increasingly sophisticated computational algorithms. Although these perspectives have substantially advanced the understanding of artificial intelligence, they do not completely resolve the distinction between information encoded within a system and information that can be functionally employed in generating intelligent behavior. This paper develops a theoretical framework centered on causal accessibility, defined as the context-dependent capacity of internally represented information to participate causally in task-relevant computational operations. The central hypothesis is that some changes in cognitive capability may arise not from the acquisition of fundamentally new representations or operations, but from changes in the functional accessibility, coordination, and compositional reuse of existing computational resources. The framework distinguishes representational availability, decodability, causal accessibility, compositional usability, and behavioral realization. It introduces a formal account of capability elicitation under fixed model parameters and specifies the conditions under which internal causal interventions could distinguish genuine functional access from mere informational correlation. Three theoretical propositions clarify why encoded information need not be behaviorally usable, why changes in accessibility can alter task performance without parameter modification, and why behavioral equivalence does not imply mechanistic equivalence. The proposed framework is situated in relation to mechanistic interpretability, causal abstraction, in-context learning, global workspace models, and theories of compositional intelligence. Its potential contribution is an integrated explanatory account of how latent computational resources may become operative cognitive capabilities. The paper does not claim to establish a novel fundamental law or provide new empirical findings. Instead, it offers a theoretically explicit and potentially falsifiable research program for investigating capability formation, emergence, and the organization of artificial intelligence.
Language: English
Open Access status: green
OpenAlex ID: W7237360338
Cites:
Cited by:
Related to: