Mind Over Model: Why AI May Never Capture the Soul of Consciousness

Table of Links

Abstract and Introduction

1. Extents and ways in which AI has been inspired by understanding of the brain

   1.1 Computational models

   1.2 Artificial Neural Networks

2. Embodiment of conscious processing: hierarchy and parallelism of nested levels of organization

3. Evolution: from brain architecture to culture

   3.1 Genetic basis and epigenetic development of the brain

   3.2 AI and evolution: consequences for artificial consciousness

4. Spontaneous activity and creativity

5. Conscious vs non-conscious processing in the brain, or res cogitans vs res extensa

6. AI consciousness and social interaction challenge rational thinking and language

Conclusion, Acknowledgments, and References

1. Extents and ways in which AI has been inspired by understanding of the brain

1.1 Computational models

Presocratic Greek philosophers already stated that any description of reality is produced by human beings (our brain) through models which necessarily displays physical limits (Changeux & Connes, 1995). This is also a logical limitation. As Kant (1781) argued, all our experiences make essential reference to our own, perforce finite, perspectives that we can never transcend, which means that we are, in a manner of speaking, prisoners of our brains (Evers, 2009). In other words, we are epistemically limited because of our finitude and the physical constraints of our brains to produce models. Accordingly any mathematical modeling, including AI that relies on and is informed by computational models, shall never be able to give an ”exhaustive” description of reality, physical or biological. The issue is thus whether and to what extent any model that assumes that a brain function/feature like conscious processing may be implemented in exactly the same way in different physical structures, either biological or artificial (i.e., functionalism), can be useful to partially or fully describe or simulate the brain (e.g., generating testable hypotheses about human consciousness, such as, for example, the Global Neuronal Workspace theory and its experimental evaluation (Mashour, P. Roelfsema, J.-P. Changeux, & S. Dehaene, 2020b)), notwithstanding the fact that, even if potentially useful (Smaldino, 2017), any biological model of the brain today is an oversimplification of neuroscientific data and of their actual biological complexity (Chirimuuta, 2024). This is not to deny that brain models may be useful and adequate even if limited in the number of details they contain, depending on the specific aspects of the brain that are modelled and on the relevant level of description. It is theoretically possible, for instance, that not all the lower-level details are necessary in order to reproduce, predict, or simulate some higher level properties, and therefore that higher levels of description of the system provide more relevant and more sufficient information (Hoel, 2017) (Rosas et al., 2020). Yet, if the goal is a comprehensive description or even a simulation of the whole brain, then any computational model would be insufficient (Farisco, Kotaleski, & Evers, 2018).

The relevance of low-level neural organisation for the simulation of conscious processing has been denied by functionalist philosophers (Butlin et al., 2023). Recently, Peter GodfreySmith has argued that the functional similarity of two systems is a matter of degree (i.e., it depends on the extent that a system needs to be understood, in coarser or finer-grained ways) (Godfrey-Smith, 2023). The crucial point is what a degree of similarity is necessary for duplicating an entity like conscious processing. Multiple realizability is the thesis that the same kinds of mental capacities can be manifested by systems with different physical architectures (Cao, 2022). This thesis has been contested. For instance, following Ned Block (Block, 1997), Rosa Cao recently argued that strict functionalism imposes quite stringent constraints on the underlying physical structure rather than eventually allowing multiple realizability. In fact, complex integrated functions (like consciousness) impose more constraints, including at fine-grained levels, than functions that can be decomposed into simpler independent functions (Cao, 2022).

Other theoretical accounts of consciousness have a somehow ambiguous critical stance towards functionalism and multiple realizability. This is the case, for instance, of the Integrated Information Theory (IIT) (Albantakis et al., 2023). IIT relates conscious processing to “integrated information” (i.e., the amount of information generated by a complex of elements, over and above the information generated by its parts). Intrinsic information is defined as what makes differences within a system. Conscious processing is ultimately identical with intrinsic information: a system is conscious if it generates information over and above its constituting parts and independently from external observers-interpreters. This is the reason why, according to IIT, “a digital simulation of the brain cannot be conscious”, neither in principle or in practice. On the other hand, a neuromorphic silicon-made computer could be conscious, because it could be composed of neuron-like elements intrinsically existing and characterized by conceptual structures (i.e., cause-effect repertoires) similar to ours (Tononi, 2015).

Therefore, IIT is against functionalism, arguing, in the spirit of Edelman (Edelman, 1992; Tononi & Edelman, 1998), that an exclusive focus on functions ignoring the physical structure cannot explain consciousness (Tononi, 2015). Particularly, re-entry processes are crucial for explaining consciousness: only systems with feedback loops can integrate information, while feed-forward systems cannot become conscious. Thus IIT is not functionalist because it stresses the crucial role of physical components substrate necessary for information generation and integration, that is, for conscious processing. Furthermore, according to IIT, a system that functions like a conscious human is conscious only if it uses the same architecture (i.e., re-entrant) as humans. Even if not functionalist, IIT eventually admits the possibility of replicating consciousness in different systems.