Large language models and the human mind

Artificial intelligence’s large language models (LLM’s) are premised on the thesis that processing language provides a promising pathway to produce thoughts. This presumption is an ontic reduction of Heidegger’s ontological thesis that thinking and language are deeply connected. Or more precisely: The core premise of AI’s large language models is an ontic reduction of Heidegger’s ontological thesis that real or genuine thinking amounts to listening to the voice of language or to what language has to say and so reveals. Large language model AI’s interest in past datasets is also an ontic reduction of Heidegger’s ontological thesis that as historical beings our past is deeply involved in our understanding, so that real thought is rooted in remembrance and thus in what tradition shows and so reveals. These ontic reductions should warn us. The human mind is not closed. It’s open towards wholly new and yet unforeseen possibilities. It doesn’t merely pay attention to past patterns to create similar patterns. A relentless adoption of ontic large language model based AI could therefore lead to a closing of the human mind.

An improved version of the semantic argument

The third premise of my semantic argument can be improved significantly. I realized this a few days ago while preparing my talk on the argument for the upcoming OZSW conference on 11 and 12 December at VU University in Amsterdam. Premise 3 of the argument can be stated as follows. If M1 and M2 are the meanings of two positive predicable generic expressions E1 and E2 such that RefSet(M1)=RefSet(M2), then M1=M2. In my paper I provide an elaborate defense of this premise. But in fact I do not need this premise. All I need for my argument to go through is the following weaker modified premise. If M1 and M2 are the meanings of two positive predicable generic expressions E1 and E2 such that RefSet(M1)=RefSet(M2), then 'x is E1 iff x is E2' is necessarily true. This modified premise is indeed weaker, since the original premise entails it but not vice versa. After all, E1 and E2 having the same meaning obviously entails that 'x is E1 iff x is E2' is necessarily true, while it is not the case that the necessary truth of 'x is E1 iff x is E2' suffices to conclude that E1 and E2 have the same meaning. Now, I suggest to replace the original premise 3 of my argument with the modified weaker premise. By doing so the conclusion of my argument, namely that there are no positive contingent universally held properties, still follows. Therefore, by doing so we have obtained a stronger version of the semantic argument, making its conclusion even more plausible.

Metaphysics and language

Let us look at a general objection that is often leveled against the method of drawing metaphysical conclusions from claims about the structure of language. The objection is that ontological consequences simply cannot be deduced from claims about linguistics, however plausible. Any attempt to slide from the lingual to the ontological plane is by its very nature not permissible.

Is this objection convincing? Now, it seems reasonable to hold that the categorical structure of language reflects, at least to some extent, the categorical structure of the world. If this is so, linguistic categories provide us with insight into the world’s structure. That is to say, a conceptual analysis of the linguistic categories of language gives us clues on the type of entities that the world contains. For example, proper names and definite descriptions reflect the ontological category of objects, general terms reflect the category of properties and relational expressions that of relations.

Thus the fact that natural language contains proper names, definite descriptions, general terms, and relational expressions, is one of the reasons we have for believing that there are objects, properties, and relations in the world. That is to say, the linguistic structure of natural language reveals certain ontological categories.

Moreover, natural language seems not only to reflect reality’s ontological categories, it also seems to reflect a number of reality’s ontological patterns. Language for example points to the ontological pattern of objects having properties. After all, in language general terms are predicated of proper names and definite descriptions. Now, if it is reasonable to hold that language mirrors both reality’s categories and its patterns, then performing conceptual analysis of linguistic structures can provide us with (defeasible) insight into the ontological categories and patterns of the world. That is, from the structure of language ontological consequences can be derived.

Further, given that semantics is a part of the conceptual analysis of language as well, it follows that semantic theses can also have ontological consequences.[2] All in all, I conclude that the generic objection is unconvincing.

Two caveats: First, by maintaining that the structure of language reflect the structure of reality, one is not committed to the radical position that we must determine the structure of reality solely by analyzing the structure of language.[1] I do not claim that we know that there are objects (properties, relations) only because we know that there are proper names (general terms, relational expressions). I do not hold that metaphysical inquiry reduces to linguistic analysis.

Second, someone who holds that the structure of language reflect the structure of reality is also not committed to the even more radical position that there are objects (properties, relations) by virtue of there being proper names (general terms, relational expressions). I do not claim that the structure of reality is ontologically dependent on the structure of natural language.

[1] Miller (2002, pp. 67-68) accepts this radical position and holds that Frege accepted it as well. This view is shared by other contemporary metaphysicians, such as Hofweber (2009), who argues that metaphysical inquiry should consist of analyzing linguistic expressions, and Thomasson (2009), who holds that metaphysical questions are to be answered by analyzing the application conditions of the terms of our language.

[2] An example would be the famous argument for fatalism (the claim that whatever will happen in the future is already unavoidable) from the principle of bivalence – a fundamental principle of semantics according to which every proposition (including those about the future) is either true or false. I do not endorse this argument, but merely mention it as an example of how philosophers use semantic principles to argue for ontological claims.

- Miller, B. (2002). The Fullness of Being. Notre Dame: University of Notre Dame Press.
- Hofweber, T. (2009). Ambitious, Yet Modest, Metaphysics. In: Chalmers, D. J., Manley, D., & Wasserman, R. (eds.), Metametaphysics: New Essays on the Foundations of Ontology (pp. 260-289). Oxford: Oxford University Press.
- Thomasson, A. L. (2009). Answerable and Unanswerable Questions. In: Chalmers, D. J. et al., Metametaphysics: New Essays on the Foundations of Ontology (pp. 444-471). Oxford: Oxford University Press.