humanflow

Large language model

Also called LLM

A large language model is a system trained on very large amounts of text to predict what comes next, and which generates prose by making that prediction repeatedly.

Last reviewed 15 August 2026 · The HumanFlow team

In plain English

It is a very elaborate autocomplete. Given everything written so far, it estimates what the next fragment of text should be, appends it, and repeats.

There is no plan and no draft. The whole document is produced one prediction at a time, left to right.

A worked example

What a model is doing at a single step, with plausible continuations ranked by how likely it considers each.

Text so far:  "The results of the study clearly"

Candidate next words, most likely first:
   demonstrate
   show
   indicate
   suggest
   contradict
   embarrassed

The top four are near-interchangeable and all appear constantly in this construction. Whichever is chosen, the sentence stays on a well-worn track.

"Contradict" is grammatical and much rarer here. "Embarrassed" is nearly impossible after "clearly".

Repeat this a few hundred times and you get text made mostly of high-ranked continuations, which is exactly the property detectors measure as low perplexity. The prediction mechanism and the detection signal are the same thing seen from two sides.

Why it matters for AI detection

Everything in this category follows from the prediction mechanism. Why generated text reads smoothly, why it repeats stock phrasing, why it invents citations that look right, and why detectors work at all are all consequences of next-token prediction rather than separate quirks.

It explains the fabrication problem in particular. A model asked for a reference produces the most probable-looking reference, and a plausible one is exactly what maximising probability yields. Nothing in the mechanism checks whether it exists.

It also explains why detection is difficult and getting harder. As models improve, their output moves closer to the distribution of ordinary human text, which shrinks the gap the detectors are measuring.

Commonly confused with

Generative AI
Generative AI covers systems that produce images, audio, video and code as well as text. A large language model is the text-and-code kind. Every LLM is generative AI; not all generative AI is an LLM.
Token
A token is the unit a model reads and writes — roughly a word fragment. The model is the thing predicting them. Model output length and cost are both counted in tokens.

Read next

Part of the AI detection glossary.