Knowing with Large Language Models: a Paradigmatic Break

This article focuses on the nature of the transformations in our relationship to knowledge induced by the massive use of large language models (LLMs). Noting the proliferation of alarmist and even catastrophic discourses since ChatGPT went online in November 2022, we argue that, to understand what can change with LLMs, we must place ourselves on the side of epistemology. The hypothesis developed is that the normative framework inherited from knowledge engineering and symbolic computing is called into question by the successes of connectionist computing. Thus, the question addressed through this article is less about whether it is possible to qualify as knowledge the texts generated by a language model, than to understand what can be known through them. What intelligibility of texts and audio and visual content, on which they are pre-trained, can LLMs claim?

Our analysis, between the disciplines of computer science, the philosophy of knowledge and information and communication sciences, first compares the functioning of LLMs, as opposed to the major principles of knowledge engineering. The fundamentally probabilistic and empirical approach of language models, associated with distributional semantics and the architecture of transformers, makes it possible to capture words and their context thanks to impressive internal representation capacities. Sometimes complementary, these techniques nevertheless oppose the symbolic approaches from which knowledge engineering emanates, which are based on a computo-symbolic conception of intelligence as well as semantics, and on a vision of knowledge stemming from formal logic and rationalist philosophies. Secondly, we show why this functioning of LLMs encourages us to observe epistemological prisms and limits to their empirical effectiveness, both of a social nature. This involves considering sociotechnical mediations and their consequences on the knowledge regime of objects generated by generative AIs. We demonstrate that it is necessary to carry out an analysis of the material and immaterial representations that guide the development of LLMs, the algorithmic choices, implicit or explicit, that frame empirical effectiveness. We are interested in those that are made during the constitution and scaling of training data corpora (for example with LAION) and during the evaluation of the performance of a LLMs, through the choice of benchmarks in training (for example, the Measuring Massive Multitask Language Understanding benchmark).

To conclude, this article opens a programmatic field of research to continue exploring the epistemological consequences of connectionist computing, a discipline that is very successful with LLMs and more generally with deep learning.