The Absence of a Realizational Level of Analysis in Contemporary AI Approaches

Description (short, index-oriented)

This paper analyzes a shared methodological limitation across contemporary artificial intelligence approaches: the absence of an explicit realizational level of analysis.

Despite substantial diversity in computational methods — including statistical learning, causal modeling, logical reasoning, interpretability techniques, grounding-oriented research, and hybrid systems — the realization of attributed cognitive effects is rarely formalized. Cognitive properties are therefore commonly inferred from behavior, relational structure, or transparency rather than from identifiable realizational carriers.

The paper demonstrates that interpretative difficulties in AI arise not from shortcomings of individual methods, but from the lack of a unified realizational perspective applicable across heterogeneous architectures. The principle of non-statistical intelligence is positioned as a methodological criterion for distinguishing computational success from realizational grounding.