# A language model does not reason 14 September 2026 · Essay · 5 min read By [Ahsan Fazal](https://axiomatic.digital/en/about#ahsan), Founder and CEO In 2025 I announced “reasoning” as a feature. The word was wrong. A language model picks the most likely next text, astonishingly well, and does not know what it does not know. So: the model writes, the sources decide what is true, code does what must be exact, the professional decides. In September 2025 I wrote in an internal update that Inora’s new platform could now reason, further than a general model. In December I announced “Structured reasoning” as a new feature. The word was wrong. Not the feature; the word. What happened was this: the model first wrote out intermediate steps, read them back, and only then wrote the answer. That helps. The answers get better. But it is not reasoning, and anyone who calls it that is selling something that is not there. I did it myself. That is why I am writing this. ## What a language model is A large language model keeps choosing the most likely next piece of text, from statistical patterns in everything it has read. That sounds small. It is not. In July 2025 I was reading how nursing knowledge is built from tasks that hang together in hidden ways, and I wrote that language models are ridiculously good at exactly that: matching statistical patterns. I still think so. A language model finds connections in language that nobody wrote down. It turns a messy note into a readable handover in seconds. It copes with a question in everyday speech, with a typo and no capitals. What it does not do is know what it does not know. It sounds just as sure when it is wrong. It likes to tell you what you want to hear; researchers call that sycophancy, and it is not the fault of one model but the behaviour of the whole kind. And it does not know who is asking, what work that person does, or which version of the protocol applies today. ## A brain is not a computer In December 2025 someone sent me an argument that algorithms could replace human judgement, because people differ so much from one another. I wrote back that such an argument stands or falls on one assumption: that a brain is a computer. That piece is [Variation is not noise](https://axiomatic.digital/en/updates/variation-is-not-noise). I do not accept it. The parallels between how people learn and how language models are trained are striking, and I find them fascinating. But similar is not the same. Nobody knows yet what a language model does to our own thinking over time. There is hardly any research that follows people over years. Anyone who is already sure that AI makes us dumber, or smarter, knows more than there is to know. I am careful with both conclusions. What I do see: the right model, used well, helps people do better work, as long as it stays a tool and does not replace the thinking. I have made the mistake myself, in the other direction. Earlier I had told the content expert I was working with that the knowledge library did not need all those books on communication: the language model could do that on its own. I was wrong. Talking fluently does not mean a model knows how to have a hard conversation with an angry family. That takes professional knowledge, from a source someone chose. ## What that means for care If a language model does not reason, you have to divide the work. This is how we design it. **The model writes.** It orders, summarises and puts what a care worker wrote into the organisation’s format. It is good at that. **The sources decide what is true.** Every answer should rest on its sources, and whoever asks should be able to see the source: its layer, its version and the passage. If the answer is not in the sources, the system should say so, say where it looked and who owns the subject. It must not guess. **Code does what must be exact.** A sum should run as code on the real data, not in the model’s text. That is how we build Inora. The headings of a report come from the organisation’s register, not from the model. The line Inora does not cross belongs in the code, not in an instruction someone can talk it out of. **The professional decides.** Inora decides nothing about a client. It supports the professional with sources and with what the care worker wrote. The decision stays with the person who answers for it. ## And in the Netherlands Two things make this urgent here. First, it is already happening. According to Nivel, just over half of the 890 care workers it surveyed used AI at work in the second half of 2025, two in three of them on their own initiative. A general model always sounds as if it knows what it is saying. Anyone who does not know what it is believes it. Second, the model changes under your feet. In August 2025, shortly after a new model came out, I saw that on launch day it had been almost a different model from what it was a few weeks later. If care rests on a language model, every new version has to be tested on the organisation’s own questions before staff work with it. You cannot leave that test blindly to another language model: it has the same urge to agree with you. A model may help check, but the measure is what the organisation’s quality staff consider a good answer. Our starting point is that what an organisation puts into Inora trains no model at all. In February 2026, when I ran our first agent that searches the sources in several steps, I wrote that it was still very dumb and ran without a system prompt. That was not modesty. It was a description. The intelligence is not in the model. It is in what you build around it: sources with an owner, code that guards the line, a test on real questions, and a professional who decides. A language model does not reason. It does not have to. It has to write well, and we have to make sure that what it writes rests on something. ## Sources - Sharma et al., [Towards Understanding Sycophancy in Language Models](https://arxiv.org/abs/2310.13548), arXiv:2310.13548. - Nivel, [Kunstmatige intelligentie (AI) in de zorg: de meningen van verpleegkundigen, verzorgenden, verpleegkundig specialisten, begeleiders en praktijkondersteuners](https://www.nivel.nl/nl/publicatie/kunstmatige-intelligentie-ai-de-zorg-de-meningen-van-verpleegkundigen-verzorgenden), 9 April 2026. ## Work in care and use Inora? Sign-in and help go through your own organisation: your team’s project lead and ambassadors can help you. ## Read on - [Variation is not noise](https://axiomatic.digital/en/updates/variation-is-not-noise.md): The case for replacing professional judgement with algorithms stands or falls with one assumption: that a brain is a computer. Where a judgement has no right answer apart from the judgement itself, difference between professionals is information. Give them the same sources, not the same conclusion. - [Rebuilt on open building blocks](https://axiomatic.digital/en/updates/rebuilt-on-open-building-blocks.md): Six weeks in, we are rebuilding Inora from the ground up on open-source building blocks. A small company does not win by building everything itself, or by letting one platform make its choices. What we take, what stays ours, and what it costs. - [“The professional decides” is a boundary, not a shield](https://axiomatic.digital/en/updates/the-professional-decides-is-a-boundary-not-a-shield.md): Reading on human oversight of AI in health care ends in one conclusion I share: a signature proves presence, not supervision. “The professional decides” has to mean two things at once. A line Inora does not cross, and a design duty to put in front of her what she needs to decide without redoing the work.