Definition
A language model is an AI system trained on text to estimate what is likely to come next given the text so far. Applied over and over, that one ability lets it write paragraphs, answer questions, summarise documents, rewrite copy and sort text into categories. "Large language model", often shortened to LLM, refers to the very large models behind today's general-purpose tools; the mechanism is the same.
The guide to how language models generate text explains the steps. The tool assembles everything in the input, the visible prompt, its own hidden instructions, the conversation so far and any attached material, into one text. The model scores possible next pieces, one is chosen with a controlled amount of randomness, and the process repeats until the answer is complete. It commits to each piece as it goes rather than planning the whole answer first.
What a language model is not matters as much. It is not a database of verified facts, it is not a search engine unless a tool adds a search step, and it has no separate process for checking its output against the world.
In practice: drafting, summarising and classifying with review
Marketing teams get most value where prediction is strong: structure, phrasing, variations, rewriting for another audience and summarising material that has been supplied. The guide to AI for keyword research shows the pattern in one field: a model is good at widening a seed list and suggesting the words customers use, and knows nothing reliable about search volumes or what the results page shows today.
The AI foundations hub builds the working routine around that division: supply the facts, ask for what prediction does well, and check the specifics. The guide to synthetic personas and simulated research applies the same reasoning to research: a model answering as a customer reproduces how such customers are written about, which is not evidence of what real customers think or do.
Why fluency and accuracy come apart
The model always produces the most plausible continuation. When the true answer is in its input, or appeared often and consistently in its training text, the plausible continuation is usually true. When it is not, the plausible continuation is something with the shape of a true answer. Company-specific facts, recent changes, niche topics and requests for sources are where this happens most, and nothing in the output marks the difference.
That is why the fixes are about input and review rather than wording. Putting the facts in front of the model, manually or through GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training., removes the need to recall them. The guide to how grounding and retrieval work explains the automatic version. Review then concentrates on figures, names, dates, quotes and claims, because a draft that reads well has said nothing about whether it is accurate.
A garden centre's plant care copy reads well and gets two months wrong
Bloomwhisker Garden Centre asked a language model to draft care notes for 40 shrubs in its spring range, giving it only the plant names. The drafts were clear, friendly and on brand. A horticulturist on the shop floor reviewed them before printing and found that 2 of the 40 gave the wrong pruning month: one told customers to prune an early flowering shrub in late winter, which would have removed that year's flowers. The model had produced the timing that is typical for shrubs in general. The team now pastes the supplier's care sheet into the prompt with an instruction to use only that and to write "check" where it is silent, and the horticulturist still signs off every card.
Related terms
GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. makes a language model answer from supplied material, which is the main defence against invented specifics. An EmbeddingA list of numbers that represents a piece of text so that texts with similar meaning sit close together, used for search, retrieval and clustering. is produced by a related kind of model that turns text into positions by meaning rather than writing it, and is used to find and group material. Fine-TuningTraining an existing language model further on a set of examples so that a style, format or task pattern is built into the model itself. trains an existing language model further on examples, shaping style and format but carrying facts only as an impression. Brand VoiceThe consistent personality a brand expresses in its words across channels, with tone adjusting to the situation and the audience. is something a model can imitate from examples and instructions; whether the result is on voice is still decided at approval.
Common confusions
- It looked it up. Unless the tool has a search or retrieval step, the model wrote from learned patterns.
- It knows our company. It knows what was common in its training text, which rarely includes a company's current products, prices or policies.
- Confidence signals accuracy. Recalled and invented facts come out in the same fluent, confident style.
- A better prompt removes errors. Clear prompts help, but specifics have to be supplied and checked.
- The same prompt gives the same answer. Choice involves controlled randomness, so prompts are judged on several runs.
