A Theory of Consciousness: Integrated Information Theory 4.0

What might separate an embodied language model from a human being? The answer most people might offer is that human beings are conscious, which is to say that they have a feeling of being themselves and have an experience associated with the words they produce or the words they take as input – something that the model does not have. But is that true? What is the basis of rejecting the model or really anything as a conscious entity? Let alone the language model – perhaps your vehicle has the same feeling of tiredness at the lower limit of its fuel tank that you have after running for long? Just because there is no verbal or behavioral access to its inner life does not necessarily mean that it lacks one.

To answer these questions is to define what consciousness is. But that is a difficult task. Effectively, one needs to start with labeling systems/states as conscious or not, then come up with a definition of consciousness by identifying the differentiating factors, which would then explain why the systems/states initially labelled as ‘conscious’ (or not) are conscious (or lack consciousness).  This essentially provides insight only into the criteria for initial labeling. Therefore, its utility for consciousness is limited to the extent that the initial labeling accurately distinguishes between conscious and non-conscious systems/states. As you can see, defining consciousness is hard to do without being circular, as consciousness is an inherently subjective phenomenon. 

Despite the challenges of consciousness research, there are several theories associated with consciousness whose different approaches are often classified as addressing the ‘easy,’ ‘hard,’ and now ‘real’ problem [1]. The ‘easy problem’ is to find how material interactions give rise to observable behaviors like perception, cognition, learning, etc. This is akin to reverse engineering an airplane; the ‘real problem’ is to find material mechanisms that can account for properties of conscious systems; the ‘hard problem’ is to explain why there is consciousness. Current neurobiological theories of consciousness fall under the broad categories of higher-order theories, global workspace theories, re-entry and predictive processing theories, and integrated information theory [2].

In an excellent thought-provoking C&R colloquium by Larissa Albantakis, Professor of Computational Psychiatry at the University of Wisconsin-Madison, held on 2025, January 15th, the speaker discussed the latest version of te Integrated Information Theory (IIT 4.0) [3], and IIT 4.0’s view of consciousness. Dr. Albantakis made it a point to clarify that IIT is not attempting to address how the physical world gives rise to consciousness (the ‘hard’ problem) but rather how experience or phenomenological properties of consciousness can be accounted for in physical terms. It defines existence, intrinsicality, information, integration, exclusion, and composition as essential properties of an experience, while the properties specific to a particular experience, like that of experience of a space or time, are defined as its accidental properties.

At its core, IIT is the idea that conscious systems have cause-effect power that is not reducible to the cause-effect power of any of its partitions or subsets. This is called a system’s integrated intrinsic information and is represented by ϕ. The estimation of ϕ requires identifying the cause-effect structure of a substrate by finding its transition probability matrix (TPM). The value of ϕ reflects the quantity of consciousness, and its shape reflects the quality. Therefore, IITs answer to the question of ‘what makes something conscious?’ is that the systems or states that have a higher value ϕ are more conscious that those that have a lower value. Please see figures 2 and 3 for IIT’s explanatory identity and methodology. The talk also mentioned IIT wiki [4], which provides extensive detail on the formalism of the theory. The fundamental identity of the theory is that every phenomenological property (essential and accidental) of the conscious experience must be accounted for by a corresponding property of the ϕ-structure. The talk briefly clarified this by explaining how IIT may account for experience of space by visualizing space as a collection of spots of different sizes and laying out the properties of the experience in terms of relations between them.

Further, IIT may predict consciousness in some systems considered non-living. According to IIT, systems with recurrent connections, but not with purely feed-forward connections, can be conscious depending on their ϕ-structure. 

IIT is one of the most mathematically formalized theories of consciousness, which means that it makes specific predictions that can, in principle, be tested empirically. It, however, faces the same problem that any theory of consciousness is bound to face – that of there being no independent objective measure of consciousness available for validation. But, in conjunction with subjective reports and behavior/response-based inferences as measures of consciousness, IIT offers great promise in advancing our ability to detect, quantify, and know the quality of consciousness in different states and systems.

References

1. Seth, A. K. (2016). The real problem. Aeon.

2. Seth, A. K., & Bayne, T. (2022). Theories of consciousness. Nature reviews neuroscience, 23(7), 439-452.

3. Albantakis, L., Barbosa, L., Findlay, G., Grasso, M., Haun, A. M., Marshall, W. et al. (2023). Integrated information theory (IIT) 4.0: formulating the properties of phenomenal existence in physical terms. PLoS computational biology, 19(10), e1011465.

4. Hendren, J., Grasso, M., Juel, B. E., & Tononi, G. (2024). Integrated Information Theory Wiki (https://www.iit.wiki). Center for Sleep & Consciousness, University of Wisconsin–Madison.