Quick answer
This site's plain-words version: an LLM is the text-writing engine inside many AI assistants; it learns patterns from huge amounts of text and writes by predicting the next word, so its output can sound right and be wrong. IBM (read October 7, 2026) calls LLMs "a category of deep learning models trained on immense amounts of data".
Key takeaways
- An LLM is the model behind a chat assistant, not the assistant. IBM, AWS, Google Cloud and NVIDIA each describe it in terms of very large amounts of data.
- Google's glossary defines a hallucination as plausible-seeming but factually incorrect output, and IBM's page lists it as a limitation of LLMs.
- An LLM is not a search engine or a database. This is this site's clarification; IBM describes retrieval as a separate method that connects a model to knowledge bases.
- This site's suggestion: let an assistant reword text you supply, and check every fact, number, name, quote and citation it adds before it leaves your hands.
- This site's seller questions: which model provider, what happens when the model is wrong, whether sources are shown, and whether your data is used for training.
What is an LLM?
This site's plain-words version: an LLM is the text-writing engine inside many AI assistants; it learns patterns from huge amounts of text and writes by predicting the next word, so its output can sound right and be wrong. IBM calls LLMs "a category of deep learning models trained on immense amounts of data" (page read October 7, 2026). Separately, IBM writes: "LLMs work as giant statistical prediction machines that repeatedly predict the next word in a sequence."
What do the vendors' own pages say an LLM is, does and gets wrong?
Four explainer pages give similar definitions and lists of uses; IBM and AWS also say LLMs can be wrong; NVIDIA's challenges section is about the cost of building them. Each cell quotes one company's page; all four companies sell AI or cloud services, so read them as descriptions, not tests.
| Vendor and page | Definition, in its words | Used for, in its words | What the page says about errors, cost or other limits |
|---|---|---|---|
| IBM, explainer article | "a category of deep learning models trained on immense amounts of data" | "summarizing an article, debugging code or drafting a legal clause" | the model generates information that is "false or misleading while sounding plausible" |
| AWS, "What is" page | "very large deep learning models that are pre-trained on vast amounts of data" | "answering questions, summarizing documents, translating languages and completing sentences" | "As impressive as they are, the current level of technology is not perfect and LLMs are not infallible. However, newer releases will have improved accuracy and enhanced capabilities as developers learn how to improve their performance while reducing bias and eliminating incorrect answers." |
| Google Cloud, LLM overview | "a statistical language model, trained on a massive amount of data" | "text generation, machine translation, text summarization, question answering, and creating chatbots" | not stated on the pages we read |
| NVIDIA, glossary entry | "deep learning algorithms that can recognize, summarize, translate, predict, and generate content using very large datasets" | "Financial advisors can use LLMs to summarize earnings calls and create transcripts of important meetings." | no sentence on errors; on cost, its challenges section says building and maintaining models "have been a barrier to entry for most enterprises" |
The cells are three different kinds of statement: IBM and AWS on wrong output, NVIDIA on the cost of building models, and for Google Cloud the page we read states none. A gap in a page is not a finding about the product.
Why does an LLM make things up?
Google's glossary defines a hallucination as "The production of plausible-seeming but factually incorrect output by a generative AI model that purports to be making an assertion about the real world." The same glossary lists confabulation as a synonym. IBM describes the same behavior as the model generating information that is "false or misleading while sounding plausible."
This site's reasoning, not a quotation: an LLM writes by predicting the next word, and predicting a likely word is a different job from confirming a fact. So it can produce fluent text that is wrong.
A worked example, as an illustration and not a test. You ask an assistant: "Write a reply telling a customer when we close on Sundays." You did not give it the hours. A reply naming a closing time reads naturally either way, so style cannot tell you whether the time came from you or from a guess. This site's caution follows: do not treat a confident tone, or the model's explanation of its own answer, as evidence.
What is an LLM not?
An LLM is not a product, a search engine or a database. All three are this site's clarification; the quotes below only illustrate.
- Not a product. IBM writes: "LLMs are easily accessible to the public through interfaces like Anthropic’s Claude, Open AI’s ChatGPT, Microsoft’s Copilot, Meta’s Llama models, and Google’s Gemini assistant, along with its BERT and PaLM models." This site's reading: IBM's list mixes products and models, so it illustrates the point without proving it. This page adds nothing about any item in it.
- Not a search engine. This site's clarification: the pages above describe a model that produces text. Looking things up is a separate step. IBM describes retrieval-augmented generation (RAG) as "a method for connecting a pretrained model with external knowledge bases."
- Not a database. This site's clarification: Google Cloud calls an LLM "a statistical language model", which is a different thing from a store of exact records. If you need an exact price or a customer's order, the record lives in your own system.
LLM versus chatbot versus AI, in this site's plain words. AI is the broad field. An LLM is one kind of model within it. A chatbot is an app people talk to; Google's glossary says "A chatbot is an application of a large language model."
Which jobs are safe to hand to an LLM and which need a check first?
Hand over wording and structure; check facts. This table is this site's suggestion, not a vendor ranking, and it assumes you give the assistant the facts to work from.
| Task | Safe to delegate? | Check first |
|---|---|---|
| Reword or shorten text you wrote | Yes, with a read-through | That your meaning and any figures survived |
| Summarize a document you paste in | Mostly | Names, dates and amounts against the original |
| Draft a customer reply from facts you supply | Yes | Anything the draft added that you did not supply |
| Answer a factual question from memory | Treat as a lead only | Every claim against a primary source |
| Quote prices, laws, citations or statistics | No | Each one at its source |
| Medical, legal or tax questions | No | This site's suggestion: use a qualified professional. Separately, IBM's page, in a note within its discussion of prompts, says "LLMs are not recommended to be used for medical advice!" |
Two-minute check, this site's suggestion: underline every name, number, date, quote and link the assistant added, and check the ones a customer, regulator or bank could hold you to against a document you trust. Delete or fix what you cannot confirm; do not ask the assistant whether it is sure.
Do I need to know this, and where will I meet an LLM without the name?
You can ignore the technical detail, but you still need the question list below. This site's observation, not a measured count: features sold as "AI" for drafting email replies, summarizing meetings or answering a website chat are places to look, and a tool may not say which model runs it. See our pieces on an AI email responder, an AI chatbot on your website and AI meeting notes.
If the tool writes to customers or touches money, the failure mode matters. For assistants that act on their own, read AI agents for small businesses; for where all this fits, see the practical guide to AI automation.
What should I ask a seller who says "built on an LLM"?
Ask which model, what happens when it is wrong, whether sources are shown and what happens to your data. All four are this site's questions; each note says what the sources do and do not cover.
- Which model provider, and can it change? A seller should be able to name the model under their product. What a vendor does if the model is swapped is theirs to answer.
- What does the tool do when the model is wrong? AWS says "LLMs are not infallible" and IBM lists hallucination as a limitation. Ask for the checks built in: a human approval step or a limit on what it can send.
- Does it answer from my documents, and are the sources shown?
- Is my data used to train the model? The explainer pages read do not address this, so the answer sits in the seller's terms. Get it in writing before you paste customer details.
How was this page made?
Definitions come from IBM, AWS, Google Cloud and NVIDIA explainer pages; the error statements from IBM, AWS and Google's machine learning glossary; NVIDIA's challenges section covers cost, not errors. This page was drafted with Claude, a model made by Anthropic; Claude and ChatGPT appear only inside IBM's list. Vendor statements are reported as theirs and were not tested here.
Next step: take one thing an assistant wrote for you this week, underline its facts and spend two minutes checking them against a source you trust.
FAQ
What does LLM stand for?
Large language model. AWS writes it as "Large language models, also known as LLMs".
Is a chat assistant the same thing as an LLM?
IBM says LLMs are accessible to the public "through interfaces" and gives a list of examples; Google's glossary says "A chatbot is an application of a large language model." This site's reading: the LLM is the engine, the product is the chat window and settings around it.
Why do LLMs hallucinate?
Google's glossary defines hallucination as "The production of plausible-seeming but factually incorrect output by a generative AI model that purports to be making an assertion about the real world." It says confabulation is a synonym. IBM's page describes the model generating information that is "false or misleading while sounding plausible". The link to next-word prediction is this site's explanation below, not a quotation.
How is an LLM different from a chatbot and from AI?
This site's plain-words reading: AI is the broad field, an LLM is one kind of model inside it, and a chatbot is an app that can sit on top of one.
Can an LLM be fixed so it never makes things up?
None of the pages read says it has been eliminated. AWS writes that "LLMs are not infallible", in a passage that also expects newer releases to be more accurate, and IBM lists hallucination among the limitations it describes. Plan for checking, not for perfection.