Are we sure we understand the meaning of “meaning” for AI?

Artificial Intelligence is not so new. Expert systems exist since the Sixties. Expert systems have never really captured the imagination of people. Internet was the first really surprising artificial intelligence that captured the imagination. The ability to write a query in a natural language and receiving data, papers photos from Internet was something really impressive.

The main change happened with Large Language Models and Generative AI. Applications such as ChatGPT are able to respond to questions, create new documents, produce images.

These applications give the impression that a computer really understand questions and comments. We often read articles that seem to corroborate the idea that computers understand meaning.

But this notion is easily misleading. We, humans, understand meaning in the sense that words and sentences represent an objective outside reality. We believe that the word “table” represents a real object and even words such as “love” or “society” have external references.

Machines understand meaning in a totally different sense. Machines adopt the “structural linguistics” theory originated by Ferdinand de Saussure that looks at languages as systems of signs not of symbols. For machines, meaning is a very large set of relationships between signs without any external reference. In particular, co-occurrences of signs in linguistic contexts. Machines adopt the structural view that meaning is “difference of meaning”.

Understanding what “meaning” means for machines is critical for the correct deployment of the current forms of AI.