Recurrent Integration and the Empirical Grounding of Phenomenal Consciousness in Artificial Intelligence Systems

Abstract

Artificial intelligence systems continue to increase in sophistication, renewing the questions of what structurally distinguishes conscious experience from computation. This paper develops a unified framework for consciousness by combining Recurrent Processing Theory (RPT) with a weakened form of Integrated Information Theory (IIT). The aim is to articulate a mechanistic account in which recurrent feedback stabilizes perceptual contents and structural integration unifies them into a single, irreducible experiential field, and then to evaluate whether contemporary AI architectures exhibit these features. Using this framework, the analysis examines the consciousness-relevant organization of two major classes of models: Large Language Models (LLMs) and Emergent Models (EMs). The discussion shows that EMs, due to their intrinsically recurrent dynamics and globally interdependent state evolution, more closely approximate the structural conditions identified by the RPT and weak IIT account than do standard feedforward transformer-based LLMs. The paper also reconsiders the debate between phenomenal and access consciousness by providing an RPT and weak IIT interpretation of the Sperling experiment and by showing how EMs offer a way to render the posited structure of phenomenal consciousness empirically tractable.