Fintan Mallory (2025), in H. Cappelen & R. Sterken (eds.), Communicating with AI: Philosophical Perspectives, Oxford University Press (chapter; Crossref print date 2026-09-29) — DOI-10.1093/9780191998317.003.0003
Key takeaways
- A trained neural network is best understood as a measuring instrument whose internal states have been altered by interaction with a dataset, and a generative model as a simulation built on that measurement. The puzzles about interpreting LLMs are therefore the familiar philosophy-of-science puzzles about instruments, models and simulations, not puzzles about other minds.
Seeded from the abstract and a skim, before the reading in NOTE-090. What follows is what the work says about itself.
Mallory argues that large language models are stochastic measuring devices. A measuring device is an artefact built so that its internal states change in response to its environment in ways that yield information about that environment. Neural networks count as stochastic because training ends in initialisation-dependent local minima rather than a global minimum. Just as a chemist puts a thermometer in a liquid, machine-learning researchers put networks into datasets to find their patterns. On this view, debates over interpreting language models and their outputs become familiar philosophy-of-science debates. Generative AI is the use of such models in simulation, and “communicating with AI” is a pattern of interaction with scientific instruments.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 0 seconds of active reading over 0 sessions. It was filed Deferred, unread. NOTE-090 is the close reading of 2026-09-26, and it placed the work: Active — worth reading as the most developed instrument-centred alternative to agentive framings of LLMs, and for its tidy division into model-interpretation, text-interpretation and linking problems; it offers an analogy, not a refutation (its own words, § Interpretation and Communication), and should be read in the published OUP version where possible.
When seeded, before the reading in NOTE-090, the entry gave a priority for a deeper reading: medium — t=0, but it is the most ML-relevant philosophy-of-science piece in the batch, and the skim captures the thesis rather than its defence.
What the seeding asked a deeper reading to check (NOTE-090 is that reading):
- It is a deflationary, instrument-centred alternative to the agentive framing of LLMs, and it bears directly on how the anthology describes what a trained model is. Borderline for home: anthology, but its question is the philosophical one of representation and interpretation.
- Check whether the “stochastic” qualifier does real work, since calibrated instruments are usually valued for repeatability, and whether seed-dependence undermines the measurement reading.
- It pairs with LIT-208 (same volume) and LIT-184/LIT-151 on models and representation.
Access when seeded: philpapers.org is behind Cloudflare. Read the full preprint (PhilPapers MALLLM.pdf, 24 pp., PDF metadata “Language models are stochastic measuring devices (revised)”) through the Wayback Machine: the abstract, all section heads (marked §), and the last two sections. The PhilArchive record page could not be reached (Wayback 404, and WebFetch 403). The earliest Wayback capture of the PDF is 2025-04-16, so published: gives April 2025 only as a month; the actual deposit date is unverified and may be earlier. Keywords are mine, not the author’s. Crossref lists the OUP chapter.