Kant and Strawson on Thought as a Cognition Through Concepts and Its Dependence on Experience, Or Why Generative AI Cannot Think (Yet)

In the Critique’s Analytic of Concepts, Kant analyzes our human capacity for thought by investigating what specific representational capacities and acts are required to realize its task of a “cognition through concepts” (A68/B93, A69/B94), by which he understands a representational, true or false relation of general concepts to individual objects. Similarly, Strawson, in the first chapter of Individuals, conceives of the representational relation to individual objects as the task of conceptual schemes, and soon arrives at Kant’s problem of individuating the objects specifically of concepts. The main point of contact between Strawson and Kant that I want to emphasize concerns the essential link between a descriptive or conceptual relation to objects in thought and a demonstrative or perceptual (intuitive) relation to the same objects in experience: According to both, the capacity of relating concepts to objects requires the possibility to relate these concepts to objects that are demonstratively related to through sense-perception. This in turn leads to the requirement to explain how experience can directly relate to the objects of our concepts. As in the case of thoughts, Kant argues that an act of combination is essential to explain how this requirement is fulfilled. In judgment, what is combined are concepts; in the synthesis of intuition, what is combined are sense-impressions and representations of homogeneous (spatial or temporal) parts. In either case, the representational relation is made possible by an act of combination, through which representations are first of all related to one another and thereby to objects and their features. This is based on what I call Kant’s Master Argument about Representation (see B129f.). At its core, it analyzes the representation of combination in objects or facts in terms of the combination of representations. As elaborated in the A-Deduction (see A96-103), the apprehension of sense-impressions requires the reproduction of homogeneous parts, which in turn requires the recognition of the object through the consciousness of one and the same act as representing one and the same object of all these qualitative and formal features. Only this achieves experience of objects. After sketching these arguments in the first half of my talk, in its second half, I begin to tease out what I take to be their direct implication: that generative AI, at least in the form of LLMs but maybe more generally, cannot be taken to think the world. I argue that this is so because, first, AI cannot relate concepts to objects of experience; and, second, because AI cannot cannot achieve unity of experience, as Richard Evans has recently claimed for the machine learning system he calls an Apperception Engine, since this engine does not and cannot – as Evans himself admits – exercise the final act of a conscious recognition, which is required for the other acts of synthesis and thus for the consciousness of objects to be possible in the first place. I end by speculating on what it would take for AI to be able to think and experience in the relevant sense.