Abstract
In this paper I approach the problem of consciousness from a non-Cartesian perspective, that is, by focusing not on a single, rational, well-informed, introspective, epistemic agents, but on a distributed system of interacting agents that need to reach knowledge about their own states.
In the first part of the paper, I introduce the "game of knowledge" (GK). GK is based on some well-known results in epistemic logic concerning knowledge reasoning in distributed systems. Basically, GK allows one to create and fully control contexts in which (a) there are some possible conditions in which some agents (we shall assume they are male prisoners) can be placed (e.g. wearing a red or a green fez) and (b) each agent involved is invited to tell (is supposed to do his best to come to know) in which condition he has been placed, without being afforded direct inspection of the condition itself (e.g. without looking at his own fez). GK is a fair and inferential game: there is no cheating (e.g. no sceptical doubts are allowed on whether it is really true that there are only green and red fezzes) and the possibility of direct, immediate, intuitive or privileged access to one's own acquired state is excluded in principle.
GK can be played with different degrees of difficulty depending on the nature of the conditions imposed on the prisoners (this simply means that agents who qualify as good players at a higher level can play successfully at any other lower level, but not necessarily vice versa). In the second part of the paper, I use different versions of GK to analyse different types of agents, and to identify them as types of players depending on the kind of version of the game they can play successfully, at least in principle. I show that some versions can be played by any group of logical agents, including artificial agents, whereas the highest version of GK can be played only by conscious agents, so it can be taken as a procedural definition of consciousness or as a criterion to identify it. The conclusion is that, currently, the only conscious agents we know of are human.
As with the Turing Test, the paper does not presuppose a taxonomy of agents. The point in using KG is not to make sure that e.g. artificial agents cannot count as good players of a certain version of the game. On the contrary, the exclusion of any appeal to direct access aims at excluding any implicit bias, whether cognitive or psychological. The attempt is to understand what kind of agents we are dealing with by seeing what kind of version of KG they can play successfully. The conclusion - only humans are conscious - is therefore not an argument in favour or against present or future AI.