The Ninety Percent

The literature that clinicians cite when they advise patients to be careful with AI is, by the field's own measurement, almost entirely AI-touched. That is not a gotcha. It is the clearest evidence yet that the two-track vocabulary of patient power was never about danger, and always about position.

The finding

In August 2026, three researchers posted a preprint estimating that 90 percent of biomedical papers published in December 2025 and archived in PubMed Central showed signs of AI-assisted writing. Across the whole of 2025 the figure was 77 percent. For 2024, 52 percent. The work has not been peer reviewed, and one of its own authors said his first reaction was that the team had made a mistake. Further checks did not overturn it.

The estimate comes from measuring how often vocabulary characteristic of large language models appears above its pre-2023 baseline. Applied to 2024 abstracts, the same method raised an earlier published estimate from 13.5 percent to 31 percent. Nothing about 2024 changed in the interval. The detector did, and that fact belongs in any honest account of the headline number.

What the number does not say

Sovereignty is a discipline of verification before it is anything else, which means handling this number with the care we would demand of anyone using it against us. It is a population estimate rather than a verdict on any single paper: it says the corpus-wide vocabulary shift is consistent with roughly nine in ten papers having had some contact with a model. It cannot tell you which ones. It cannot separate a fabricated results section from a researcher polishing the grammar of a second language, and the study's own finding that use runs highest among authors in countries where English is not dominant suggests a large share of the signal is translation and polish, which is the most defensible use of the tool there is.

One figure inside the study does earn alarm on its own terms. An estimated 58 percent of results sections in December 2025 papers showed signs of AI, against 78 percent of discussion sections. The results section is the part of a paper that is supposed to be nothing but observation, and it is the place where a model's tendency to fabricate would do the most damage. That is the number to worry about, and it is not the one in the headline.

The asymmetry

Set the two cases side by side. A researcher uses a language model to draft, translate, or tighten a paper, and the field calls it writing assistance and asks for a disclosure line. A patient uses the same model to make sense of her own pathology report, and the field calls it a risk: something clinicians should screen for, something to raise with the doctor before she acts on it.

The act is the same act, asking a machine for help with dense technical language that carries consequences. The difference is not risk, either, and the direction of the risk runs the other way. A hallucinated sentence in the published literature propagates into reviews, guidelines, and every downstream reader, including the clinician who will use it to advise the patient. One patient's bad summary reaches one patient. If the field were allocating caution by how far an error travels, the warnings would be pointed at the journals.

What differs is standing. One party is presumed competent to use a tool and disclose it. The other is presumed to need supervision. That is a claim about who holds authority, dressed as a claim about safety.

Why researchers never had to ask

Notice how the professional norm actually arrived. No body granted researchers permission to use language models. They used them, at scale, and the institutions wrote disclosure policy afterward to describe what was already true. Nobody proposed that researchers stop, and nobody invented a program to make them AI-ready. The capability came first and the vocabulary followed it.

The vocabulary offered to patients runs in the opposite direction. Informed, educated, engaged, activated, empowered: every term in the lineage of granted words presumes an institution standing above the patient, holding the thing to be handed down. Researchers were never asked to become engaged with AI. They simply used it, and the word for what they did was borrowed from ordinary professional life, where competence is assumed.

Patients are doing the same thing at the same moment for the same reason, and about one in three adults already do it. This is the fourth of the four tests in plain view: a vocabulary that can describe a profession's adoption of AI as routine, while describing a patient's identical adoption as a hazard, is not describing the technology at all. It is describing who was already assumed to be in charge.

See also The Engagement Frame Wrote the AI Agenda and Objections, Taken Seriously.