In short A large language model (LLM) is an AI system trained on enormous amounts of text that works by predicting the next word, over and over, to produce fluent language. It’s the engine behind tools like ChatGPT, Copilot, and Gemini. LLMs are brilliant at drafting, summarizing, and explaining — but they generate plausible text, which isn’t always accurate, and they don’t ’look things up’ unless connected to a tool that does.

Every AI chatbot you’ve heard of — ChatGPT, Microsoft Copilot, Google Gemini, Claude — runs on something called a large language model, or LLM. The term gets used constantly and explained rarely. So here’s what an LLM actually is, in plain English, and why understanding its one core mechanic tells you exactly what it’s good and bad at.

Intelligent machines, broadlyArtificial IntelligenceLearns from dataMachine LearningLayered neural networksDeep LearningCreates contentGenerative AIPredicts the next word to produce languageLLM

An LLM sits several layers deep inside AI — Hover any part for detail.

Where LLMs sit in the AI family

Quick orientation first. Artificial intelligence (AI) is the broad field. Inside it is machine learning (ML) — systems that learn patterns from examples instead of following hand-written rules. Inside that is generative AI, which creates new content. LLMs are the specific kind of generative AI built for language. So an LLM is a type of generative AI, which is a type of machine learning, which is a type of AI. (Our AI, ML & GenAI module has a diagram of this nesting, and the What Is Generative AI? article covers the wider family.)

The one thing an LLM actually does

Here’s the mechanic that explains everything: an LLM predicts the next word.

That’s it. Given some text, it predicts the most likely next chunk of text (technically a “token,” roughly a word-piece), adds it, then predicts the next, and the next. Given “The capital of France is,” it predicts “Paris.” Given the start of an email, it predicts a plausible continuation. Do this thousands of times in sequence and you get fluent paragraphs, working code, or a structured report.

The “large” part refers to two things: the model was trained on a vast amount of text (much of the public internet, books, code) and it has a huge number of internal parameters — the dials it tunes to make those predictions. That scale is what lets a “next word predictor” produce writing that’s often indistinguishable from a knowledgeable human’s.

Key point: an LLM doesn’t retrieve stored answers or “know facts” the way a database does. It generates the most plausible-sounding next words based on patterns in its training. That’s why it’s fluent — and also why it’s sometimes confidently wrong.

What this explains about its strengths

Because it’s a language-generation engine, an LLM shines at anything that’s fundamentally about producing or transforming text:

  • Drafting — emails, documents, proposals, job descriptions.
  • Summarizing — turning long reports or threads into short ones.
  • Rewriting and explaining — changing tone, simplifying jargon, translating.
  • Coding — writing, explaining, and debugging code.
  • Brainstorming — generating options to react to.

If the task is “make some language out of this,” an LLM is often superb. Getting good results is mostly about giving clear instructions — the skill of prompting, covered in the prompts, chatbots & copilots module.

What this explains about its weaknesses

The same mechanic explains the failures. Because an LLM produces plausible text rather than verified text:

  • It hallucinates. It can invent facts, citations, statistics, or quotes and present them with total confidence. Fluency is not accuracy.
  • It doesn’t inherently know your data. An LLM learned general language from public text — it knows nothing about your company’s internal policies or numbers unless you provide them in the prompt (or connect it to your data via a technique like RAG).
  • It has a knowledge cutoff. Unless connected to live tools, it only “knows” what was in its training data up to a certain date.
  • It’s not a calculator or database. It predicts text, so it can fumble precise arithmetic or exact lookups unless it’s using a tool.

The practical rule: use an LLM to draft, but verify anything that matters. This “trust but verify” discipline is the heart of the AI risk & human review module.

How to use the term

  • “That chatbot runs on an LLM, so it won’t know our internal policy unless we give it to it.” (understanding the limits)
  • “It’s an LLM — great first draft, but fact-check the numbers.” (trust but verify)
  • “The model hallucinated that citation; it sounded real but doesn’t exist.” (naming the failure)
  • “We connected the LLM to our docs so it answers from our actual data.” (RAG in plain terms)

An LLM isn’t magic and it isn’t a search engine — it’s a remarkably capable next-word predictor trained at enormous scale. Hold that one idea and both its brilliance and its blind spots make perfect sense. For the broader picture, see What Is Generative AI? and AI vs Machine Learning.

Related reading: What Is Generative AI? · AI vs Machine Learning · AI Agents at Work