Technology
What a language model does, what tokens and a context window are, why a model invents things and what it does and does not remember from your input.
An LLM, or large language model, is a computer model trained on enormous amounts of text, allowing it to predict which word follows a sequence of words. Repeating this process produces sentences, answers and summaries. It is the engine behind assistants such as ChatGPT, Claude and Mistral Vibe.
Most people wondering what an LLM is have already used one. The abbreviation appears in quotations and policy documents alongside terms such as generative AI, tokens and hallucinations. This article explains those words for someone who is not a programmer, then moves to the question that matters at work: what happens to what you enter?
The technical explanation is simplified. Provider agreements and legislation are linked to sources, with a reference date of 13 October 2026. The practical overview of chatbots and personal data is on ChatGPT, Claude, Mistral Vibe and the GDPR.
A large language model receives text and calculates the best-fitting continuation. It appends that piece, then starts again. An answer emerges word by word.
Anthropic describes it the same way in its glossary: the underlying model is trained to predict the next word given the preceding text.
To predict well, a model must have picked up many patterns: grammar, contract structure, usual time limits. Those patterns reside in billions of numbers, the parameters. Nobody put them there as rules.
The result resembles understanding. But the model has no list of facts to consult. It has a sense of what is probable. That difference explains both its strengths and its mistakes.
Generative AI is the umbrella term for AI creating new content: text, images, sound, video or code. An LLM is the type creating text. Every LLM is therefore generative AI, but not all generative AI is an LLM: an image-generating model works differently. Newer models often process images or sound too, and are called multimodal.
Training happens in stages. First, the model receives an enormous text collection and endlessly practises one task: predicting the continuation. Each time it gets it wrong, the parameters are adjusted slightly.
Then comes adjustment so the model answers questions and follows instructions. Fine-tuning is the name for an additional training round on extra data, for example to focus a model on one subject or writing style.
When you ask a question, no training takes place. The parameters are fixed and the model calculates an answer from your text. Afterwards, it is exactly the same model as before.
The model therefore does not learn from your conversation while you have it. Whether the conversation is later used to train another version is a different question, discussed below.
Because training stops at a particular point, a model knows nothing about what happened afterwards. That point is called the knowledge cutoff. If an assistant still gives a current answer, the product around it has looked something up.
A model reads tokens, not words and letters. In its explanation of tokens, OpenAI says a token may be a character, part of a word, a whole word or punctuation. For English text, that page gives a rule of thumb of about four characters, or three-quarters of a word, per token. The ratio differs in other languages and between models.
The context window is all the text the model can consider for one answer. In its documentation on context windows, Anthropic calls it working memory, distinct from the text used to train the model. Everything counts: product instructions, the entire conversation so far, added documents and the answer being written.
Windows have grown substantially in recent years. As of October 2026, the same page lists windows of up to a million tokens, depending on the model. Bigger is not automatically better. Anthropic writes that accuracy and recall decline as more tokens enter the window. Three hundred pages for one question do not automatically produce a better answer than the three pages that matter.
A model generates the continuation that best fits the preceding text. Usually that is also correct. Sometimes it is a sentence sounding like a fact that is not one: a judgement that does not exist, an article number slightly wrong, a source with a plausible title. This is called a hallucination.
In its guide to reducing hallucinations, Anthropic writes that even the most advanced language models can produce text that is factually incorrect or inconsistent with the supplied context. It is inherent in how a model works.
ChatGPT is not an LLM. It is an OpenAI product containing an LLM. The same applies to Anthropic's Claude and Mistral Vibe (formerly Le Chat).
Almost every privacy question concerns the product and provider, not the model. The same model may be offered under different terms: in a free app, a business workspace or another supplier's software. Ask who runs the model and under which agreements too.
A closed model runs with its maker or a party the maker designates. You use it through an app or programming interface. Your input therefore always goes to someone else.
With an open model, the maker publishes the parameters, often called weights. Anyone with sufficient computing power can run the model themselves, on their own equipment or with a supplier of their choice. For example, Mistral's model overview publishes open-weight models under an open-source licence alongside models available only through its own service. Read the licence: “open” does not mean the same thing for every model.
An open model you run yourself does not send input to its maker. The work shifts to you: security, management and equipment. An open model used through someone else's service is no different for your data from a closed model. That service's terms then apply.
The answer has three layers.
The model “knows” what you said earlier because the product sends the entire conversation again. This is not model memory. Start a new conversation and the window is empty.
Conversations are stored with the provider for its retention period or until you delete them. If the product has a memory feature, it also stores notes about you. Under the GDPR, this is ordinary processing of personal data.
Whether conversations are used to train a later model version is the provider's choice. It is not a model characteristic. On its Enterprise privacy page, OpenAI writes that training uses, among other things, data from versions of ChatGPT for individual users, while business product data is not used by default.
This is explained by plan in ChatGPT Free, Plus, Business or Enterprise: which plan can you use for business? and Claude AI and the GDPR. Business options are compared in ChatGPT alternatives for business use.
A model does not store training text as an archive, but may have memorised portions. In its opinion on AI models of 18 December 2024, the EDPB wrote that model anonymity must be assessed case by case. Among other things, it must be very unlikely that someone can extract personal data through questions. That is one reason to disable training on input for business use.
| Term | What it is | Why it matters |
|---|---|---|
| LLM (large language model) | A text-predicting model trained on enormous amounts of text | The engine behind every AI assistant |
| Generative AI | AI creating new text, images, sound or code | The broader category containing LLMs |
| Parameters or weights | Numbers in which the model stores patterns | Published for open models, allowing you to run them yourself |
| Training | Adjusting the model on large amounts of text | Your input may end up here if the provider uses it |
| Use (inference) | Having the model calculate an answer | Does not change the model |
| Token | The unit in which a model reads and writes text | Determines how much text fits and its cost |
| Context window | All text the model considers for one answer | What falls outside does not exist for the model |
| Hallucination | An answer sounding plausible but incorrect | A reason to request and check sources |
| Open model | A model with published weights | Running it yourself keeps input with you; the licence determines permitted use |
| Model and product | The model calculates; the product stores, connects and sets terms | Privacy agreements concern the product and provider |
The AI Act does not use LLM. It refers to general-purpose AI models. Article 3, point 63, defines such a model as an AI model displaying significant generality, capable of competently performing a wide range of distinct tasks and able to be integrated into various systems and applications. Large language models fall within that definition. The chatbot you open is legally an AI system built on such a model.
Article 53 places obligations on the model provider. It must maintain technical documentation, provide information to parties incorporating the model into their own systems, have a copyright policy and publish a summary of training content. Models under free and open-source licences are exempt from the first two duties, unless they pose systemic risk.
The most powerful models receive additional obligations. Article 51 categorises a model as posing systemic risk if it has high-impact capabilities, presumed above a threshold for computing power used in training.
According to the European Commission guidelines for these model providers, these rules have applied since 2 August 2025. The Commission has been able to enforce since 2 August 2026, and models already on the market before 2 August 2025 must meet the requirements by 2 August 2027.
Using an assistant at work does not make you a model provider. Article 53 obligations are not yours. The obligations that do apply are explained in The AI Act explained. The GDPR also applies to everything you enter.
PrivaSearch is neither an LLM nor an assistant. It addresses the third step above: giving the model only what is necessary, without names.
Documents are analysed inside PrivaSearch. A language model running within PrivaSearch and Dutch patterns, such as the checksum for a BSN (Dutch citizen service number), find personal data. No external AI provider participates in this step. Detected data is replaced with pseudonyms. Because no model detects everything, a colleague checks the result alongside the original and approves it. Only then is a document searchable. See How PrivaSearch works.
An assistant can then search through a connection using the open MCP standard. This works with Claude, ChatGPT Business and Mistral Vibe. The assistant receives no file in its context window, but a limited number of released, pseudonymised passages with a short source reference.
What the assistant receives becomes part of the conversation with its provider, and what you type goes there directly. A business account with a data processing agreement therefore remains necessary. Pseudonymisation is not anonymisation: for your own organisation it remains personal data. More about the rules is on The GDPR and the AI Act.
To see what a model does and does not receive, book a demo.
An LLM is a computer model that has learned how language works by processing vast amounts of text. It receives text and predicts what follows, a small piece at a time. This produces answers, summaries and translations. The abbreviation stands for large language model. It is the technology behind assistants such as ChatGPT, Claude and Mistral Vibe.
Generative AI is the broad category: all AI making new content, such as text, images, music or video. An LLM is the part working with language. A model creating only images is generative AI but not an LLM. In practice, the terms overlap, partly because newer language models can process images and sound.
Not exactly. ChatGPT is an OpenAI product in which a large language model performs the calculations. The product adds a window, conversation history, tools and terms. The model is regularly replaced with a newer one while the ChatGPT name stays. For privacy, the product matters most: the plan determines what happens to your input.
The model itself does not change through your conversation. Within one conversation, it “knows” what you said because the entire conversation is sent again each time. The provider does store conversations according to its retention period. Whether they are later used to train a new model depends on the provider and plan. Business plans usually do not do this by default.
Because it generates probable text rather than looking up facts. Usually the most probable continuation is also correct. Sometimes it produces a source, number or quotation sounding right but nonexistent. Such an error is a hallucination. You reduce the likelihood by giving the model source text and requesting quotations. You cannot eliminate it.
Yes. The AI Act calls these general-purpose AI models. Obligations such as technical documentation and a summary of training content rest with the model's maker and have applied since 2 August 2025. Those merely using an assistant do not have those obligations. Other parts of the law apply to users, and the GDPR applies to what you enter.
How retrieval-augmented generation works, how it differs from uploading and fine-tuning, why it reduces errors and which privacy questions to ask.
What the Model Context Protocol is, how a connection differs from uploading a file, what an assistant receives and which risks you should check.
What an AI detector measures, why its result is a probability rather than proof, and what to arrange before challenging a pupil or employee about it.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →