Les LLM et leur potentiel explicatif

Mots clés

Valeur explicative, Compréhension, Narratifs Explicatifs, Pragmatique, LLM, Formats d'explication

Détail de l'offre

LLMs and Their Potential for Explanatory Value

This research proposal aims to analyze the extent to which (a) the development of large language models (LLMs) renews existing discussions and questions concerning the nature of explainability and understanding, and (b) the emergence of LLMs raises novel questions and problems related to explainability in both scientific and everyday contexts.

Over the past decades, many philosophers of science have moved away from the view that explanations should primarily be understood as linguistic entities or arguments. Nevertheless, several central aspects of explainability remain closely connected to linguistic dimensions, both at the epistemological and pragmatic levels (that is, with respect to how explanations are discovered, presented, and understood, as well as how agents reason about them). Explanatory knowledge is typically communicated to both lay and expert audiences through textual means, narratives, or argumentative structures. In the formal sciences (notably mathematics), the objects to be explained and understood—such as proofs, programs, and algorithms—also possess an essential linguistic or symbolic dimension.

In this context, the development and application of LLMs can be expected to enhance the methods by which humans acquire, articulate, and reason about explanations, while also generating new challenges concerning these notions.

The proposed PhD project aims to connect scientific research on LLMs and their applications, philosophical work on AI-based explanations and understanding, and broader philosophical discussions of explanation and understanding that conceive of explanations as symbolic or linguistic entities.

A non-exhaustive set of research questions to be addressed includes:

- How can LLMs contribute to the production and analysis of explanations, for instance, by generating appropriate explanatory narratives?
- What specific epistemological problems does the understanding of LLM as an object raise?
- Do LLMs prompt a reconceptualization of explanation as primarily pragmatic and/or linguistic, rather than causal and truth-tracking?
- Can LLMs contribute to the understanding of scientific practices or results that are often regarded as resistant to understanding, such as simulations or other AI-based methods?
- Do LLMs challenge traditional epistemological accounts that associate understanding with mental representation or the grasp of reasons?
- What role do LLMs play within the broader framework of explainable artificial intelligence (XAI)?
- What specific epistemic roles might LLM-based methods play in the development of explanatory understanding?
- Are there particular epistemic norms associated with the use of LLMs when they are intended to convey understanding?
- When LLMs are employed to foster understanding, are their uses and outputs associated with specific explanatory values?
- Can LLM-based methods contribute to all forms of explanation and understanding or just to specific non-causal ones?
- Are the potential explanatory gains associated with LLM methods domain-specific?
- What forms of spurious explanation or illusory understanding are specifically associated with LLM-based methods, and which risks warrant particular attention?
- Is “understanding with an LLM” a distinct epistemic category, different from individual or collective understanding?

The PhD project will address a selected subset of these questions. The anticipated stages of the research are as follows:
- A survey of cases in which LLMs appear to contribute to explanatory understanding
- An analysis of the relevant philosophical literature on explanation and understanding
- Cross-comparative analysis of the selected cases
- A synthesis of insights derived from the case studies and the philosophical literature
- Conclusions concerning how LLM-based methods may contribute to explanatory understanding
- How do LLM methods from other AI-based methods on the above issues?