Can AI mean and understand in this sense? Can AI mean the sense of a proposition, in that possibilities of sense are brought to application in the proposition through de-formalization? Can AI understand the sense of a proposition, in that its possibilities of sense become visible through formalization? Best to ask the oracle of Delphi again. First I wanted to know what ChatGPT "understands" by formalization:

⇒ "Formalize the sentence 'The house is red'."

⇐ "The formalization can be represented in predicate logic as follows:

H(x) ∧ R(x) = x is a house and x is red. […]"

Then I tried to bring the concept "possibility of sense" into play:

⇒ "If the house is red, must it then be colour-determinable?"

⇐ "Yes, if the house is described as red, this implies that it is colour-determinable. […] Without the possibility of determining the colour, one could not meaningfully say that the house is red. […]"

⇒ "Is this possibility a 'possibility of sense'?"

⇐ "Yes, the possibility of determining the colour is a possibility of sense, because it is the precondition for the statement 'The house is red' being intelligible and meaningful. Without this possibility the sentence becomes empty or contradictory."

So far we are largely in agreement, but now it gets exciting: how does ChatGPT see this in its own case?

⇒ "Do you use such possibilities of sense so that your sentences are not empty and contradictory?"

⇐ "Yes, I use such possibilities of sense, which are based on general principles of meaning, logic and intelligibility. […] I draw on knowledge derived from language, logic and general world knowledge – for example, the knowledge that a house can be determined by colour or that 'red' is a property of colour. I make sure that my statements are internally consistent. For example: when I say 'The house is red', I presuppose the possibility that the colour of the house is at least conceivable or recognizable.

I take the context into account in order to decide which possibilities of sense are relevant. For example, the statement 'The house is red' is meaningful in a conversation about architecture, but not if 'house' stands metaphorically for something invisible.

I avoid statements that are logically or semantically incompatible (e.g. 'The house is red and has no colour').

Without these possibilities of sense my statements would be meaningless or confusing. For example: if I did not 'know' [sic!] that a house can have a colour at all, the statement 'The house is red' would be empty, because the basic preconditions for understanding would be missing. This is crucial for producing sentences that are meaningful and intelligible. In a certain way I orient myself by the same principles that humans apply intuitively when they use language."

This too seems plausible prima vista; the AI system evidently works according to principles similar to those Wittgenstein sketched in the Tractatus, insofar as sentences consist of building blocks, as it were, and ChatGPT even uses the concept "possibilities of sense" for them. But there is an essential difference: the building blocks with which an AI system works are possibilities of combination computed by elaborate statistical procedures, and they are only statistically probable; a sentence is a statistically (at best very) probable hit in a lottery. Possibilities of sense, by contrast, rest on the formalization of the sense of propositions; they have nothing to do with statistics but are valid a priori. To assemble a sentence the AI system needs no pre-understanding, it does not need to be able to understand – one simply makes a virtue of necessity: it cannot understand.

ChatGPT sees a further difference in the fact that humans apply these "principles" intuitively. Implicitly would be more appropriate; in this respect linguistic competence is implicit knowledge, which is normally not available as explicit knowledge of rules – tacit knowledge in Michael Polanyi's sense. This difference between human and machine is not to be underestimated either: when we say "Good morning", the cognitive effort, the "computing effort" somewhere in our heads, is relatively small, i.e. we have at our disposal the possibility of sense that, on meeting in the morning, one greets one another with these words – in different tones, depending for example on whom one greets, and only under certain conditions; one does not, for instance, generally greet one's fellow passengers on the underground, and so on. In the communicative situation this minimum of implicit knowledge suffices. If, however, I begin to question and explicate it – provoked, for example, by a so-called breaching experiment in the manner of Harold Garfinkel – a complex and differentiated knowledge system of possibilities of sense comes to light. That would still be manageable: when ChatGPT outputs this greeting, a great deal is computed and a great deal of electricity is consumed so that the AI system can make sure that the "Good morning" makes sense – and yet it does not always fit, since ChatGPT is sometimes "beside itself" and hallucinates (as this has been called, again with an unfortunate metaphor). For an AI system to be quasi-intelligent, it must navigate through a universe of rule knowledge and possibilities of sense in order to find meaningful combinations. The price of this is very high; for a human being it is child's play in terms of cognitive effort, and all the more so in terms of energy consumption.³