The basics · Layer 4

What is an LLM?

LLM (large language model)

A program trained on vast amounts of text that writes by predicting, step by step, the most likely continuation.

A way to picture itPicture a reader who has worked through an entire library without taking notes: they discuss anything fluently, but quote from memory, and that memory stops on the day they closed the last book.

The diagram · LLM

01Text corpusBooks, websites, code, collected up to a date

02Training phaseThe model tunes billions of parameters

CheckApproved data only

03Your promptQuestion, context and documents supplied

04Text predictionThe most likely continuation, step by step

05Written answerFluent text, still to be checked

How it works

01During training, the model reads vast amounts of text and adjusts billions of parameters to guess the next fragment. It picks up patterns of language, reasoning and fact, without filing them away as a database would.
02Once trained, it is frozen: its knowledge stops at a cut-off date, and it knows nothing about your internal documents unless they are supplied to it.
03For each question, it builds its answer fragment by fragment, each time choosing a likely continuation. The result is fluent and plausible, but nothing in this mechanism guarantees that it is correct.

In a business function

Procurement department

A procurement team uses an LLM to compare three maintenance bids of around forty pages each. Within minutes, the model produces a table of differences in price, lead times and penalties. The buyer checks every line against the bids before the award committee meets. The time saved on reading is real; the decision stays human.

The compliance angle

Since 2 August 2025, providers of general-purpose models must document them, put a copyright compliance policy in place and publish a summary of their training data (EU · AI Act, art. 53). Your organisation remains responsible for what it sends to the model: a data processing agreement must govern any personal data shared with the provider (EU · GDPR, art. 28). A model trained on personal data is not anonymous by default; this is assessed case by case (EU · EDPB, Opinion 28/2024). An assistant open to the public must be designed so that people know they are dealing with an AI (EU · AI Act, art. 50).

Limits to know

Its knowledge stops at its training cut-off and does not cover your internal documents.
It can state falsehoods with confidence: this is a property of the mechanism, not an occasional glitch.
Two identical questions can receive two different answers, which makes checking harder.

The question to ask your supplier

“Which model do you use, in which version, where is it hosted, and are our data used to train it, yes or no?”

Submit a brief

Where this concept sits

01Business questions
02Definitions and owners
03Connected data
04AI
05Rights and human sign-off
06Evidence

Related concepts

Read on the blog

AI governanceAI governance: five proofs of control

Frequently asked questions

Are ChatGPT, Claude and Le Chat LLMs?

They are assistants built on LLMs: GPT models for ChatGPT, Claude models for Claude, Mistral AI’s models for Le Chat. The assistant wraps the model in an interface, tools and rules.

Does an LLM understand what it writes?

It handles the patterns of language remarkably well, which often produces useful reasoning. But it does not test its statements against reality: verification has to be organised around it.

What does “GPT” stand for?

Generative Pre-trained Transformer: a generative, pre-trained model based on the “transformer” architecture published by Google researchers in 2017. Most current LLMs rely on this architecture.

Educational explainer, not legal advice. Up to date as of 29 October 2026.

What is an LLM?

LLM (large language model)

A program trained on vast amounts of text that writes by predicting, step by step, the most likely continuation.

How it works

01During training, the model reads vast amounts of text and adjusts billions of parameters to guess the next fragment. It picks up patterns of language, reasoning and fact, without filing them away as a database would.

02Once trained, it is frozen: its knowledge stops at a cut-off date, and it knows nothing about your internal documents unless they are supplied to it.

03For each question, it builds its answer fragment by fragment, each time choosing a likely continuation. The result is fluent and plausible, but nothing in this mechanism guarantees that it is correct.

In a business function

Procurement department

The compliance angle

Frequently asked questions

Are ChatGPT, Claude and Le Chat LLMs?

They are assistants built on LLMs: GPT models for ChatGPT, Claude models for Claude, Mistral AI’s models for Le Chat. The assistant wraps the model in an interface, tools and rules.

Does an LLM understand what it writes?

It handles the patterns of language remarkably well, which often produces useful reasoning. But it does not test its statements against reality: verification has to be organised around it.

What does “GPT” stand for?

Generative Pre-trained Transformer: a generative, pre-trained model based on the “transformer” architecture published by Google researchers in 2017. Most current LLMs rely on this architecture.