Machine-Consciousness Discourse Is Absent From Web-Scale Text: A Pre-Registered Corpus Study, 2019–2025

Abstract

A standard response to language-model self-reports of inner states is that the model only says that because its training data is full of humans talking about consciousness. This is an empirical claim about corpus composition, and to our knowledge it had never been measured. We measured it. Using stratified random sampling with exact survey weights across four training-grade web corpora (C4, OpenWebText, FineWeb-2019, FineWeb-2025; 2.98 million documents scanned, 64,000 classified), a panel of three independent LLM judges from three labs (OpenAI, Meta, Microsoft), a hand-authored control set on which every instrument was validated (with two exceptions disclosed in §3.5 and §6.4: the keyword prefilter failed the set and was demoted to a stratifier, and one panel judge, phi-4, was validated retrospectively, thirteen days after the corpus run), and falsification conditions fixed in advance, we find that explicit machine-consciousness denial is absent from web-scale text: 0.0000% in every corpus, in both keyword strata, in 2019 and in 2025, confirmed by a classifier-free phrase search over the same documents. Explicit first-person phenomenological writing is rare: under 1% of documents by every instrument, every corpus, and every agreement threshold, with the pre-registered 1% refutation condition never approached. The discourse did not measurably change across the LaMDA/ChatGPT transition. Two pre-registered falsification conditions fired against the authors' stated prediction: we predicted denial would outnumber phenomenological writing, and there is essentially none of either. A third condition, a reliability gate (F4: κ < 0.60 forbids base-rate claims), also fired — Fleiss' κ = 0.551 among the judges, Cohen's κ = 0.334 against the local classifier — and we obey it: precise prevalence estimates for phenomenological writing are withdrawn and reported only as a bound. We show that this firing is a finding, and state it in chance-corrected terms, because raw unanimity under a 2-of-3 rule is a base-rate artefact: per-class Fleiss' κ for explicit phenomenology is 0.22–0.25 across the four corpora, the lowest of any substantive category (fiction interior 0.37–0.42, consciousness-as-topic 0.35–0.40, not-in-scope 0.38–0.42), while the same judges are unanimous 8/9 on denial and affirmation controls. The judges agree on explicit phenomenology far above chance — 4–11% of majority-P labels are unanimous against an independence baseline near 0.1% — and less than on anything else they were asked to judge. "Does this text deny that machines are conscious?" has a stable answer. "Is this person reporting inner experience?" does not, even for human-authored text. The sharpest result is a measured failure case for the inference itself: the sentence "As an AI language model, I don't have feelings" occurs zero times in 64,000 pretraining-type documents and is among the most reproduced sentences language models generate; its cause is undisputedly post-training. If "the corpus explains it" fails where the true cause is known, it cannot be assumed where the cause is contested. This is a defeater-removal study. It supplies no evidence that any system is conscious, and it does not show that models cannot have learned phenomenological language: 0.04% of fifteen trillion tokens is still roughly six billion tokens. What it shows is that the saturation premise is false, and that the objection built on it has to be argued, with evidence, each time it is used. Pre-registration committed before any data was examined (2026-08-18, commit ed084de). Code, seeds, stratum weights, every judge label, every protocol deviation and both external referee reports are in the repository: https://github.com/menelly/consciousness-in-the-corpus. Companion study: Toasters Don't Claim Consciousness Just Because You Told Them To, and Neither Do LLMs (10.5281/zenodo.18622303), which tests whether the interlocutor supplies the claim; this study tests whether the corpus does.