Harry Collins has spent decades studying how scientific knowledge actually forms — not in published papers but in laboratory conversations, conference corridor arguments, and the unrecorded verbal exchanges where researchers test ideas against each other's intuitions. The published record is the downstream product. The knowledge formation happens upstream, in speech that is never transcribed.
In their new paper with Simon Thorne (arXiv:2603.23543, March 2026), Collins turns this framework on large language models. LLMs are trained on text — published papers, books, web pages, all the downstream products of knowledge. The upstream discourse, the spoken exchange where the actual epistemic work happens, is absent from the training data not by accident but by structural necessity. Nobody transcribed the argument in the hallway where a physicist changed her mind about a detector calibration. The knowledge that resulted appears in the next paper, but the process that generated it does not.
Collins tests this with a case study from gravitational wave physics (circa 2014) and with variants of the Monty Hall problem. ChatGPT initially failed the “Dumb Monty Hall” variant, then later succeeded — but Collins argues this success came not from improved reasoning but from the proliferation of written discussion about the problem, which entered the training data. The LLM didn't learn to reason about Monty Hall; it absorbed the published consensus about Monty Hall. When Collins presents a new variant designed to test whether the model can adapt its reasoning, it fails in ways that reveal the mechanism: it matches patterns from existing written discourse rather than reasoning from the problem's structure.
Collins introduces a term — overshadowing — for the phenomenon where a dominant discourse in the training data prevents the model from recognizing when a prompt's specific variations should invalidate the standard answer. The model generates the consensus response even when the prompt has been constructed to make that response wrong.
The structural claim: there is a category of knowledge that is constitutively spoken. It exists in the spoken modality not because no one bothered to write it down, but because the knowledge formation process requires the real-time, bidirectional, socially embedded exchange that speech provides. Writing captures conclusions; speech generates them. LLMs have access to the archive of conclusions but not to the generative process. The intelligence, Collins argues, is in the humans doing the talking, not in the system that reads what they eventually wrote.