Keyword

Large language models

9 writings

A Polyphonic Conception of AI Understanding

With Pierre Beckmann.

Argues that whether large language models understand what they talk about is a question both unavoidable and ill-framed: unavoidable because users must tell outputs that merit epistemic trust from those that do not, and no purely mathematical or statistical vocabulary can draw that line without reintroducing the concept of understanding in all but name; ill-framed because the inherited concept presupposes a single locus of understanding. Mechanistic interpretability instead reveals LLMs to be deeply polyphonic systems whose outputs emerge from parallel mechanisms of uneven reliability. Drawing on evidence of medical, mathematical, logical, and spatial understanding in LLMs, the paper engineers a conception fit for such systems: a model understands a domain insofar as it contains sound circuitry that tracks relevant structure and reliably recruits that circuitry to control its outputs—which turns seemingly idle metaphysical disputes into tractable empirical questions that can guide the allocation of trust.

AI, large language models, understanding, mechanistic interpretability, conceptual engineering, conceptual adaptation, explainable AI

Download PDF

Evolving Minds, Evolving Language: Metaphor as a Process of Conceptual Adaptation to Artificial Intelligence

Manuscript (+Slideshow)

Argues that cognitively loaded terms such as “understanding” and “reasoning,” when applied to large language models, are best understood neither literally nor as merely loose talk, but as metaphors that adapt our conceptual repertoire to new circumstances under pressure of conceptual needs. Drawing on the Strawson-Kant tradition on imagination and concept-application, inferentialist semantics, Bermudez’s theory of rational framing, and Gentner’s structure-mapping account of analogy, it develops a four-part framework: inferential transfer rather than referential transfer, the aptic normativity of conceptual needs, the empirical evaluation of transferred inferences through mechanistic interpretability, and the career of AI metaphors from novelty toward possible conventionalization. The result is a shift away from all-or-nothing disputes about the literal applicability of cognitive terms to AI toward graded and empirically tractable questions about which inferential transfers are apt, revealing why our cognitive vocabulary for AI is evaluatively and practically consequential.

AI, large language models, understanding, conceptual adaptation, aptic normativity, mechanistic interpretability

View presentation

Forms of Understanding in LLMs: A Crisis for Comprehension?

In The Routledge Handbook of Philosophy of AI. Leonard Dung, Guido Löhr, and Chris Ranalli (eds.). London: Routledge. Forthcoming. With Pierre Beckmann.

AI, large language models, understanding

Mechanistic Indicators of Understanding in Large Language Models

Philosophical Studies. 2026. With Pierre Beckmann. doi:10.1007/s11098-026-02513-1

Draws on detailed technical evidence from research on mechanistic interpretability (MI) to argue that while LLMs differ profoundly from human cognition, they do more than tally up word co-occurrences: they form internal structures that are fruitfully compared to different forms of human understanding, such as conceptual, factual, and principled understanding. We synthesize MI’s most relevant findings to date while embedding them within an integrative theoretical framework for thinking about understanding in LLMs. As the phenomenon of “parallel mechanisms” shows, however, the differences between LLMs and human cognition are as philosophically fruitful to consider as the similarities.

AI, large language models, mechanistic interpretability, understanding, world models, explainable AI

Download PDF

World Modeling in Transformers

arXiv 2609.21748. 2026. With Pierre Beckmann and André Freitas. doi:10.48550/arXiv.2609.21748

Shows that behavioral failures can make a transformer appear to lack a world model even when it has learned faithful representations of its environment. In TaxiGPT, a transformer trained on random walks through Manhattan whose failures had been read as evidence of an incoherent internal map, mechanistic analysis and causal interventions reveal a model that represents intersections and streets, tracks its position, and navigates with a goal compass; its failures stem from interference between superposed intersection features, whose consequences are limited by an “affordance packing” that groups intersections with the same legal moves. Proposes mechanistic indicators for comparing models, which show world-modeling capacities emerging at different stages of training, and argues for a shift from asking whether a model has a world model to mechanistically studying its world modeling.

AI, large language models, world models, mechanistic interpretability

Download PDF

Can AI Rely on the Systematicity of Truth? The Challenge of Modelling Normative Domains

Philosophy & Technology 38 (34): 1–27. 2025. doi:10.1007/s13347-025-00864-x

Argues that the asystematicity of normative domains, stemming from the plurality, incompatibility, and incommensurability of values, poses a challenge to AI’s ability to comprehensively model these domains and underscores the indispensable role of human agency in practical deliberation.

AI, large language models, systematicity, truth, value conflict, AI ethics, deliberation

Download PDF