What is an LLM and what AI really cannot do: a plain guide for business owners

According to Anthropic's documentation, the current Claude models have reliable knowledge up to June 2026, and OpenAI's GPT-6 Astra up to 30 April 2026, so without web search they cannot tell you anything reliable about what happened later. What is an LLM? A large language model is a program trained on a huge amount of text that predicts how a text should continue. It does not check facts on its own, which is why it can be wrong with the same confidence it shows when it is right. For a business the rule is simple: let AI draft the text, and let a person check the numbers, names and legal points.

This guide explains without technical jargon how a language model is built, what tokens, the context window and the knowledge cutoff mean, why AI sometimes makes things up and how to check its answer. If your company already applies AI to specific processes, the next step is our pillar on which business processes are worth automating with AI. If an AI tool built your website, you will find practical checks in the article vibe-coded website: a first-aid checklist, and an independent look at the code is what our technical website audit offers.

What is an LLM and how is it built: from predicting a word to an assistant

An LLM is a model trained on a large amount of text that first learns to predict the next word and is then fine-tuned into an assistant that answers questions.

In its glossary, Anthropic describes LLMs as language models with many parameters, trained on vast amounts of text data. They can write human-like text, answer questions and summarise information. For a business owner, two stages of the process matter, because each one explains a different kind of mistake.

Pretraining: the model learns how text continues

In the first stage the model processes a large corpus of unlabelled text and learns to predict the next word from the preceding context. According to the same glossary, a model pretrained this way is not good at answering questions or following instructions on its own. It can continue a text, and at this stage that is where its abilities end.

Fine-tuning: the text machine becomes an assistant

The second stage is fine-tuning on additional data and a method called RLHF. In RLHF, people rank two or more example answers and the model learns to prefer the ones that ranked higher. Only then does it become an assistant that answers questions in a chat. According to Anthropic, Claude is built exactly this way: on a large language model that was fine-tuned and trained with RLHF.

The practical consequence is this. On its own, the model writes the text that, based on learned patterns, best continues your prompt. If you need a fact from AI, give it the material to take the fact from, and check that the answer refers to it.

Tokens, context window and knowledge cutoff: three terms that explain most surprises

The model processes text in tokens, sees only what fits into its context window, and knows only what was in its texts before its knowledge cutoff.

Tokens: the unit for both text and price

A token is the smallest unit a model works with. It can be a whole word, part of a word, a single character or a byte. According to Google's Gemini API documentation, for Gemini models one token is about 4 characters and 100 tokens equal about 60 to 80 English words. Anthropic states that on earlier Claude models a token represents about 3.5 English characters, and that models from Claude 4.7 onwards use a newer tokenizer that produces about 30 percent more tokens for the same text. According to Anthropic, the number of characters per token depends on the model and on the language, so these English figures do not automatically apply to other languages.

Tokens decide two things. When you pay through the API, the price is counted in them: Google writes that the cost of a Gemini API call is determined in part by the number of input and output tokens. Tokens also measure the model's capacity, meaning how much text it can handle at once.

Context window: the model's working memory

The context window is all the text the model can see while it writes an answer, including the answer itself. In its documentation on context windows, Anthropic calls it the model's working memory and distinguishes it from the data the model was trained on. The current Claude models have a window of 1 million tokens, which according to Anthropic is roughly 555,000 words. According to OpenAI, GPT-6 Astra has a window of 1.05 million tokens.

A large window does not mean the model uses everything in it equally well. According to the same documentation, accuracy and recall degrade as the token count grows. Anthropic calls this context rot. Chat interfaces such as claude.ai can also manage the window on a rolling basis in a long conversation, so the oldest parts give way to new ones. An instruction from the start of a long conversation may therefore no longer be followed at its end.

Knowledge cutoff: the line the model cannot see past

Every model has a date up to which its knowledge is most extensive and reliable. For Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 it is June 2026, according to Anthropic's models overview. For GPT-6 Astra it is 30 April 2026, according to OpenAI's models page. The model answers correctly about newer events, law changes or price lists only when someone gives it that information: you, by pasting in the material, or the application, by searching the web.

Term	What it is	What you do about it
Token	A word, part of a word or character the model counts with; capacity and API price are measured in tokens	For long documents and API use, count in tokens, not pages
Context window	All the text the model sees when answering, including the answer	Repeat an important instruction, open a new conversation for a new task
Knowledge cutoff	The date up to which the model's knowledge is most reliable	For current topics, supply a source or turn on search
Temperature and randomness	A setting for how varied answers are; the same prompt can give a different answer	Generate an important answer twice and compare

When an AI answer disappoints you, first work out which of these limits you hit. The fix depends on it: you add the missing material, restart an overloaded conversation, or check an outdated figure against its source.

Why AI answers with confidence even when it is wrong

A hallucination is text that is factually incorrect or inconsistent with the given material, and according to Anthropic even the most advanced models produce them.

According to Anthropic's guide to reducing hallucinations, even the most advanced language models, Claude included, sometimes generate text that is factually incorrect or inconsistent with the context they were given. The same guide says the recommended techniques reduce hallucinations significantly but do not eliminate them. Critical information always needs to be validated, especially for high-stakes decisions.

The cause fits how the model is built. A pretrained model learned to continue text, and checking whether something is true was not part of that training. A fluent, confident sentence therefore says nothing about whether the answer is correct.

Same question, different answer

In its text generation guide, OpenAI writes that the content a model generates is non-deterministic and that even different snapshots of models in the same family can produce different results. Anthropic adds that even with temperature set to 0, results are not fully identical. Temperature is the randomness setting: a higher value gives more varied wording, a lower one sticks to the most probable phrasing.

For a business this matters for anything repetitive. If AI sorts invoices or replies to customers, the same input may not produce the same output every time. Such a process needs a check or a fixed rule outside the model, which our article on automation for small businesses with n8n and Make also covers. Without them, an error in one of the runs may only surface with a customer.

What AI really cannot do: five limits to plan for

A language model knows nothing after its knowledge cutoff, cannot see your company data until you give it, does not guarantee the same answer twice, loses accuracy on long texts and cannot check by itself whether it is right.

1. It does not know what happened after its cutoff. Supply current prices, laws and news yourself, or use web search and open the source of the answer.
2. It cannot see your company data. The model works with what it learned in training and with what is in the context window. RAG, a technique Anthropic describes in its glossary, puts relevant documents into the window at the time of the question. Only then can the answer rest on your material.
3. It does not guarantee the same answer. Output is non-deterministic, so for repeated tasks check a sample of the results.
4. It loses accuracy on long texts. As the token count grows, accuracy and recall of details degrade. Split a long document into parts or ask the model for word-for-word quotes.
5. It does not check itself. Techniques reduce hallucinations but do not remove them. The final check stays with a person.

Which limit will hurt you depends on the task. The table below is our recommendation, not a vendor rule:

Task	Suitable for an LLM?	What a person checks
First draft of an email or post	Yes	Tone, names and every promise to a customer
Summary of a document you pasted in	Yes, with quotes	That the quotes are really in the document
Current prices, laws and deadlines	Only with search or your own material	The source and the date it applies to
Repeated sorting of enquiries or invoices	Yes, with regular checks	A random sample of the outputs
Legal or tax decision	Only to prepare questions	An expert who is accountable for the answer

How to check an AI answer in five steps

Allow the model to say "I don't know", restrict it to your material, ask for word-for-word quotes first on long documents, have every claim backed up and generate an important answer twice.

All five steps come from Anthropic's guide to reducing hallucinations, and someone who does not know the subject can run them too:

1. Allow it to say "I don't know". According to Anthropic, explicitly allowing the model to admit uncertainty can drastically reduce false information. Add this to your prompt: "If it is not in the material, say that you don't know."
2. Restrict it to your material. Tell it to use only information from the documents you provided and not its general knowledge.
3. Ask for quotes first on long documents. For documents over 20,000 tokens, Anthropic recommends having the model extract word-for-word quotes first and only then perform the task.
4. Ask for a source for every claim. The model should find a supporting quote for each claim and retract any claim without one. Then look up at least one quote in the original document yourself.
5. Generate an important answer twice. If two answers to the same prompt differ on something essential, Anthropic treats that as a sign of a possible hallucination.

Even this routine does not remove hallucinations completely. Numbers, names, dates and anything that goes to a customer, an authority or into a contract should therefore be read before sending by the person who is responsible for the result.

In practice: how we work with language models at Denva

On this blog AI writes the first draft, and every number in an article must have a source we opened on the day of writing.

This article was also written with the help of AI, as the note below it says. The rules are the same ones we recommend to you: every number, date and product name links to official documentation that we actually opened while writing, and whatever cannot be verified does not go into the text. Model knowledge cutoffs and context window sizes change with every new version, so we state the month they apply to. We recommend the same rule for the texts on your own website. How AI answers in search are changing website traffic is covered in our article on Google AI Mode and its impact on your website.

Websites built with AI tools such as Cursor, Lovable, Bolt or v0 follow the same pattern as text. The result can look functional while its weaknesses sit below the surface. A model can write code that looks correct just as it can write a sentence that sounds correct. That is why for such websites we run a technical audit focused on the patterns typical of vibe coding.

Vendors give token conversions for English, and Anthropic says plainly that the number of characters per token depends on the language. If you plan to run documents in another language through the API, check capacity and price on a sample of your own text.

Conclusion: when to trust AI and when to check it

If you use AI for drafts of texts, emails and summaries based on your own material, carry on and add the five checking steps from this article. If you expect current facts, prices or legal answers from it, give it a source or turn on search, and open that source. If you want it inside a process that runs without a person, first decide who checks a sample of the outputs and how often. And if AI wrote your website or its code, it needs the same check as text, only from someone who knows what should be under the surface.

FAQ

Frequently asked questions.

What is an LLM in simple terms?

An LLM is a large language model: a program trained on a huge amount of text that predicts how a text should continue. Once fine-tuned into an assistant, it can answer questions, summarise a document or draft a text. It does not check facts on its own.

Is Claude or ChatGPT the same as an LLM?

They are applications built on language models. Anthropic describes Claude as an assistant based on a large language model that was fine-tuned and trained with RLHF, and OpenAI's documentation compares generating text through its API to using ChatGPT.

Why does AI sometimes make things up?

The model learned to continue text, and checking truth was not part of that training, so it can write a confident but wrong answer. Anthropic calls this a hallucination and says techniques such as allowing the model to say it does not know, or quoting the source material, reduce it significantly but do not eliminate it.

Does AI know what is happening today?

On its own it only knows what was in its texts before its knowledge cutoff. For the current Claude models that is June 2026 according to Anthropic, and for GPT-6 Astra it is 30 April 2026 according to OpenAI (as of October 2026). Newer information has to come from you or from web search.

What is a token and why does it matter?

A token is a word, part of a word or a character that the model works with. For Gemini models, Google puts it at about 4 characters of English text. Tokens measure how much text the model can handle at once and, when you pay through the API, the price.