Home / What a language model is, without the jargon
Predicting the next word, nothing more
A language model does only one thing: predict the most probable fragment of text to follow what it has already read, and a made-up answer comes out of the same calculation as an accurate one, with no internal signal telling them apart.
Claude does not know the answer to a question: it calculates which fragment of text has the best chance of following the one it has just read. This sentence is enough to describe the whole mechanism. It reads the text received, estimates a probability for each fragment likely to come next, keeps one, then repeats the same calculation with that fragment added to the text. Nothing else happens inside.
Why a false answer looks like a true one
This mechanism explains a phenomenon that Claude's help centre acknowledges as a known limitation of the product: Claude can produce made-up information and display citations that seem authoritative without being grounded in a real fact. The reason lies in the calculation itself. A plausible fragment and an exact fragment are, for this calculation, two candidates of the same kind: nothing in the process gives special treatment to verified information. A made-up phone number and a real one are written with the same syntax, so with a comparable probability in the model's eyes.
What this changes in how you read an answer
A fluent, well-worded answer is therefore not a sign of accuracy: the fluency comes from the same calculation as any potential error. One example makes this point concrete.
Question: what is the customer service phone number?
Answer: You can reach them on 01 42 00 00 00, Monday to Friday.
Nothing in the form of this answer tells you whether this number comes from a real source or from a plausible fragment produced by the same calculation as the rest of the sentence. Checking this kind of information through an independent channel settles the matter; the confidence conveyed by the wording is not enough as proof.
The next word, not the reasoning
This description also holds for long answers and detailed reasoning: each step is itself a fragment predicted from what precedes it, including from the steps the model has just written. A smoothly flowing chain of reasoning is therefore not a guarantee of accuracy at each step, it is a sequence of fragments plausible relative to one another. The next lesson on the token and the window spells out what the model actually sees to produce each of these fragments.
The prediction loop, and its possible exit into a plausible but false answer
A user asks Claude for a company's customer service phone number. Claude replies with a number written in the same confident style as the rest of its answer.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: The answer was produced with the same apparent confidence whether the number is accurate or made up, since Claude produces the most probable fragment without distinguishing a verified fact from a merely plausible one.
What this does not establish: Whether the number provided is correct or incorrect, which a direct check with the company would settle.
The three most common miscalibrations
- Too broad Claude made up this number and deliberately misled the user.
- Too narrow Nothing can be said until you have seen the exact number given in the answer.
- Beside the point This situation shows that it is better to ask Claude to cite its source for this kind of question.
- Claude produces the most probable fragment of text to follow what it has already read, without consulting a separate fact base that would validate the answer.
- Made-up information and an accurate answer come out of the same probability calculation, which explains why a false answer can look just as confident as a true one.
- Claude's help centre acknowledges made-up information and ungrounded citations as known limitations of the product, not as isolated accidents.
- A smoothly flowing chain of reasoning is not proof of accuracy at each step, since each step is itself a predicted fragment.
Open a conversation with Claude, ask a precise factual question whose exact answer you already know, then write in one sentence, in your own words, what Claude actually did to produce its answer.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Claude is providing incorrect or misleading responses. What's going on? consultée le 2026-09-02