HENUSSE Making your company readable by AI. Outside, and in.

Understand · The language model

How does a language model work? What ChatGPT actually does with your question

A language model such as the one that powers ChatGPT does not consult a directory of companies to answer you. It produces its answer piece by piece, drawing each piece from among the most probable continuations of the text, based on what it learnt during training. That is why it can be very useful and, in the same sentence, be completely confident while being wrong.

This page explains the mechanism, without jargon or shortcuts. Every claim is either a basic mechanism stated as such, or a point documented by the providers themselves or by published research, with the source at the end of the page.

What does the machine actually do when I ask a question?

Your question does not enter the machine as words. It is first divided into small pieces called tokens, or tokens in English. A token can be a whole word, part of a word or even a single character. Anthropic says that for its models and for English, one token represents about 3.5 characters. The documentation specifies that this figure varies by language, so we do not transpose it to French.

The model receives this sequence of tokens and produces only one thing at a time: the next token, drawn from among the most probable continuations. Then it starts again, taking into account the token it has just written, and starts again once more. A three-hundred-word answer is the result of several hundred of these decisions in sequence. Anthropic’s documentation describes exactly this mechanism for the initial training phase: the model is pre-trained to predict the next word from the text that comes before it in the document.

This ability rests on work published in 2017 by eight researchers in a paper entitled “Attention Is All You Need”. That paper introduced the transformer, an architecture built around an attention mechanism that enables the model to weight the relative importance of each part of the text before it. OpenAI’s GPT-4 technical report, published in 2023, describes the model as transformer-based and pre-trained to predict the next token in a document.

What the model has “before it” has a name: the context window. Anthropic defines it as all the text the model can consult to produce its answer, including that answer, and specifies that it is working memory, separate from the large corpus on which the model was trained. The model cannot consult anything outside this window: it has only what its training left behind, without being able to read or verify it again.

How a language model produces its answer A diagram in three stages. One, the question is divided into tokens, represented by labels of unequal widths: a token can be a whole word, part of a word or a symbol. Two, to choose the next token, the model has several possible continuations, represented by bars of different lengths. The longest is selected, but it is not the only possibility. Three, the operation starts again, and the text produced grows one step at a time. None of these stages verifies anything. FIGURE 1 · WHAT THE MACHINE DOES WITH YOUR QUESTION 01 Your question is divided into tokens Which school of Fren ch in Brit tany ? A token can be a whole word, part of a word or a symbol. A token is not a word. 02 It chooses the next token Brittany the most probable Finistère Morbihan The selected token is drawn from among these continuations. It is not the only possibility. 03 Then it starts again step 1 step 2 step 3 step 4 Each step takes into account the token that has just been written. A three- hundred-word answer chains together several hundred of these decisions. None of these stages verifies anything. The model composes a plausible continuation; a plausible continuation and a true continuation are not the same thing.

Why are these models called “generative”?

Because the answer is produced, not retrieved. There is no record prepared in advance somewhere for the machine to fetch and show you. Every sentence you read was composed at that moment, token by token, for you.

You can easily observe the consequence yourself: ask the same open question twice, in two new conversations, and the wording may change from one time to the next. This is not a defect. It is the nature of the process. Anthropic documents the parameter that governs this variability, temperature, and specifies that even when set to its minimum, it does not make the results completely deterministic: identical inputs can produce different outputs from one call to the next.

The same question twice, two answers: what does that change for a company?

An AI answer is a draw, not a ranking. It is composed when the question is asked, it may differ when the question is asked again, and it exists nowhere as a stable list that could be consulted, bought or corrected.

Three consequences follow for a company. A citation obtained once proves nothing by itself: it becomes information only if the same question is asked again and every trial is dated. An absence observed once proves nothing either, for the same reason. And a figure announced without the exact question, date and number of trials cannot be verified. That is the purpose of rules 03 and 07 of our measurement charter.

One question remains for every business leader: does anyone control what the machine says about my company? There is no record to correct. A provider trains its model and sets boundaries around its answers, but it does not decide what the model will say about your company when the next question is asked. Anthropic documents for its own models that identical inputs can produce different outputs from one call to the next. No provider can therefore guarantee that you will be cited. What it can commit to is the work and the measurement.

Why can the machine state false things with confidence?

Because nothing in the mechanism described above verifies anything. The model composes a plausible continuation of the text. A plausible continuation and a true continuation are not the same thing, and the model has no internal means of distinguishing between them.

There is a second reason, less intuitive and better documented than people think. A team that included OpenAI researchers published a paper in September 2025 entitled “Why Language Models Hallucinate”. Its thesis fits into one line: training and evaluation procedures reward guessing rather than acknowledging uncertainty. A model evaluated like a candidate in a multiple-choice examination has an incentive to guess rather than leave the box blank. Its confident tone is therefore not a surface accident. It is encouraged by the way these systems are measured.

The industry’s response to this problem is grounding: have the model search the web, then compose its answer from the pages actually retrieved, with links. OpenAI describes its web search tool as enabling models to access up-to-date information and provide answers with sourced citations, with inline citations included by default. Anthropic describes the same principle for Claude and specifies that source citations are always enabled for web search.

This is where the subject stops being theoretical for a company: when the machine searches, it can use only what it finds. We publish the method we use to measure these answers in our measurement charter.

Searching is not generating: the two paths of an answer Two columns are compared. On the left, without search: your question, then the model composes using what it retained during training, then an answer that displays no source. On the right, with search: your question, then the model decides to search, then pages are retrieved and read, then an answer that displays links to those pages. In both cases the answer is composed. Only the second can be opened and cross-checked. FIGURE 2 · SEARCHING IS NOT GENERATING WITHOUT SEARCH Your question The machine composes using what it retained during training An answer with no source to open WITH SEARCH Your question The machine decides to search, according to the question asked Pages are retrieved, then read An answer with links to the pages displayed In both cases, the answer is composed. Only the second can be opened, read and cross-checked by you. OpenAI and Anthropic document this decision to search and the display of links; documentation accessed on 16 August 2026.

What does a language model not do?

Four misunderstandings often arise when discussing these models with a business leader. Here are the four misunderstandings, each followed by what the model actually does.

  • It does not contain a database of companies. There is no “your company” record stored in the model that could be corrected. What the model retained about your activities, if it retained anything, is distributed across its parameters and cannot be consulted.
  • It does not systematically go to the web. Search is a tool, not a reflex. Anthropic writes that Claude decides whether to search according to the request, and answers directly, without searching, when the question concerns stable knowledge. The same documentation lists current prices and information about organisations or products among the cases in which the model searches, and established facts among the cases in which it answers directly.
  • It does not reason about your case as an adviser would. There is nobody behind the answer with an opinion about your company, a preference for you or your competitor, or an awareness of being right or wrong at the moment it writes.
  • It does not know current events by default. Its training stops on a date. Anthropic presents web search as the way to answer with up-to-date information beyond that cut-off date. Without search, something new from last week is invisible to it.

Remembering these four limitations is enough to rule out most of the unverifiable promises heard about the subject.

Frequently asked questions

Will ChatGPT read my website to answer?

Only if it triggers a search, and only if your page is among the results returned by that search. When it does not search, it answers from memory, and no page is consulted or cited.

Why do two people get different answers to the same question?

Because the answer is generated each time. Anthropic specifies that even at zero temperature, identical inputs can produce different outputs. The account, conversation memory and location add further differences.

Does the model “know” my company?

The question is badly framed. There is no record to consult. There are probabilities of text learnt during training and, if the model searches, pages retrieved when it produces the answer.

What is the context window, in one sentence?

It is the model’s working memory: all the text it can consult to produce its answer, separate from the corpus on which it was trained.

Does what this page describes apply to Claude, Gemini or Perplexity?

The underlying mechanism, dividing text into tokens and then predicting the next one, is common to current language models. We cite only documents that we have checked, published by OpenAI, Anthropic or researchers, and we give no figure concerning the other providers.

Sources

All these addresses were accessed and checked on 16 August 2026.

Page published on 16 August 2026; updated when the facts change.

WhatsApp