Table 1.

Comparisons between four perspectives on aspects of consciousness: FEP-AI, IIT, GNWT, and IWMT.

	FEP-AI	IIT	GNWT	IWMT
Levels of analysis emphasized	Functional, algorithmic, and implementational	Phenomenological and implementational	Functional and implementational	Phenomenological, functional, algorithmic, and implementational
Emphasizes either phenomenal or access consciousness	Both	Phenomenal	Access	Both
Emphasizes either intrinsic or extrinsic perspectives	Both	Intrinsic	Extrinsic	Both
Neural substrates of consciousness	A distributed pattern of effective connectivity (entailing Bayesian beliefs) across a multi-level deep temporal hierarchy, primarily generated by L5 pyramidal neurons and thalamic relays	A maximal nexus of self-cause–effect power, likely centered on posterior cortices	A global workspace realized by re-entrant connectivity between frontal and posterior cortices	Agreement with FEP-AI, except these distributed patterns are hypothesized to be integrated via the formation of self-organizing harmonic modes, so promoting communication through coherence Agreement with IIT with respect to basic phenomenal consciousness, but with specific emphasis on posterior-medial cortices as a basis for egocentric perspective Agreement with GNWT with respect to access consciousness, but with phenomenality being generated from posterior loci
Minimally conscious system	Any generative model with temporal depth and counterfactual richness; e.g., all deep belief hierarchies capable of adaptive active inference	Any system capable of generating irreducible cause–effect power over itself; e.g., a single elementary particle	Any system capable of implementing a global workspace; e.g., a computer program with a blackboard architecture	Any process capable of generating a world model with spatial, temporal, and causal coherence with respect to the system and its causal inter-relations with its environment; e.g., all mammals, possibly all vertebrates, and possibly insects
Can a system without dynamics be conscious?	No	Yes, if it is part of a configuration capable of constraining likely past and future states	No	No
Could an artificial intelligence (AI) implemented on a von Neumann architecture be conscious?	Yes	No	Yes	Probably
Is either physical or a richly structured virtual embodiment required for consciousness?	Yes	No	No	Yes
Associated concepts from machine learning and AI	Variational autoencoders Forney factor graphs with marginal message passing	Direct implementation on neuromorphic hardware capable of recurrent dynamics	Blackboard architectures	Folded variational autoencoders with recurrent dynamics Turbo codes
Are human-equivalent intelligent zombies feasible?	No comment	Yes	No comment	Theoretically conceivable, but practically infeasible

FEP-AI, Free Energy Principle and Active Inference framework; IIT, Integrated Information Theory; GNWT, Global Neuronal Workspace Theory; IWMT, Integrated World Modeling Theory.