I wanted to add the following discussion of information by Anthony Chemero, in which he contrasts an enactivist view of it from representational, computational modeling approaches like that typically seen in predictive processing.
“Predictive processing models, and the allied theoretical machinery, bounce back and forth between thermodynamic and information-theoretic understandings of information and entropy. These understandings are not identical or even equivalent, even though they are theoretically related. But the differences are rarely acknowledged by predictive processing proponents. Here is an example from Clark himself, in a recent paper whose main ideas we otherwise endorse:
‘ This is where FEP [the free energy principle] gets invited onto the stage. FEP states that living organisms that persist must minimize free energy in their exchanges with the environment. The ‘free energy' in question here is an information-theoretic isomorph of thermodynamic free energy, which is a measure of the energy available to do useful work. Useful work, in the information-theoretic story, involves fitting a model to a domain, so reducing information-theoretic free energy is improving the model.’ (Clark 2017)
This is cringeworthy, because it runs roughshod over exactly the point where care is needed. As Deacon (2012) points out, there is no analogue of work in information theory. More important for now is that neither thermodynamic nor information-theoretic understandings of information involve semantics, whereas semantics is the key point of the ecological information in dynamic enactive models. Ecological information—the information available to a moving animal in the environment—is inherently semantic because it specifies the affordances of that environment, what the animal can do in that environment, and generates and supports expectations for what that moving animal will experience as it moves. Ecological information reveals the world as significant for a given creature.
In contrast, it rarely crosses anyone's mind that thermodynamic information might be semantic; and many people forget that information-theoretic (Shannon) information is meaningless in itself, because it is generally discussed in the context of a known decoder that transforms the encoded patterns in the transmitted signal into something meaningful. But this way of framing the relationship between an organism and its environment, and the nature of the signals it receives from that environment—in terms of a transition from thermodynamic free-energy to information-theoretic free-energy—is far from theoretically benign. It is this switch that makes free energy minimization seem to have representational implications, and appear to necessitate the various accoutrement of computational theory of mind.
Why is this? Because the most compelling case for a world-representing inner-model-building brain begins by ignoring the richness of ecological information. This leads to the claim that perception offers insufficient information to guide action, and we therefore need a model to do so (Hohwy 2013). The assumption that organisms work with thermodynamic and Shannon (i.e, non-semantic) information builds that poverty in from the start. This is problematic because some (in our view) fairly unpalatable philosophical conclusions can be drawn from this particular well: if our behavior is driven directly by a model, and only indirectly by the world, then the thought that we have only indirect epistemic access to that world can begin to seem compelling (Clark 2013; 2016b; Hohwy 2013; 2016). And if we only have indirect epistemic access to the world, then isn't perception a form of controlled hallucination, a sensory veil cutting us off from the very world we appear to inhabit?”