Performance of the model-first Active Inference strategy in artificial systems

Assess how effectively the model-first approach that derives common computational properties of consciousness from Active Inference models performs when applied to challenging or controversial cases—specifically artificial systems—in inferring the existence and contents of conscious states.

Background

The paper outlines a model-first approach to theory development for the minimal Active Inference theory of consciousness (AI-C), which builds computational models of paradigmatic phenomena and compares their properties to identify common computational features of conscious versus unconscious processing. This strategy aims to minimize reliance on introspective assumptions and enable theory construction across diverse paradigms.

The authors explicitly note uncertainty about how well this approach generalizes to more challenging or controversial cases, notably artificial systems. Establishing the effectiveness of the strategy in such domains is essential for evaluating whether common computational properties can be delineated and used to infer the existence and contents of conscious states beyond human and animal systems.