AI terms in plain language.

Software that does tasks that normally need human thinking, such as understanding text, recognising images or making a prediction. AI does not learn this from rules someone writes down, but from examples.

See also: Machine learning, Generative AI

AI that does not just answer, but takes steps by itself to reach a goal: looking something up, opening a file, filling in a form or calling another program. The more an agent is allowed to do, the more oversight and limits matter.

Example: An agent that reads incoming invoices, checks them and prepares them in your accounting software.

See also: Human in the loop, MCP (Model Context Protocol), Workflow and automation

A program you talk to in plain language. Modern chatbots run on a language model. A customer service chatbot often only knows the information of that one company.

See also: Language model (LLM), RAG (search, then answer)

AI that works alongside you inside a program you already use, such as your email, word processor or code editor. You stay in control. Microsoft Copilot and GitHub Copilot are well-known examples.

See also: AI agent

AI that makes something new: text, images, sound, video or code. ChatGPT, Claude, Gemini and Midjourney are examples.

Example: Having a first draft of a quote written, or ten variations of a product photo made.

See also: Language model (LLM), Image generator

AI that creates images from a description, such as Midjourney, DALL-E or Stable Diffusion. There are also generators for video, music and voices.

See also: Generative AI, Deepfake

A large AI model trained on a huge amount of text, so it can read and write text. LLM stands for large language model. It keeps predicting the most likely continuation, and that way it can answer questions, summarise, translate and write code.

See also: Token, Hallucination, Reasoning model

The trained AI system itself: the 'brain' you call on. There are large and small models, fast and thorough ones, for text, images or both.

See also: Training, Multimodal

A model that handles more kinds of input or output than just text, for example photos, PDFs, sound or video.

Example: Turning a photo of a whiteboard into a tidy task list.

Knowing what AI can and cannot do, how to work with it sensibly and where the risks are. Under the European AI Act, organisations that use AI must make sure their people know enough about it.

See also: AI Act

How much text a model can take in at once: your question, the attachments and the conversation so far. What does not fit, the model 'forgets'. It is measured in tokens.

See also: Token

When an AI model confidently states something that is not true, such as a source that does not exist or a wrong number. It happens because the model predicts what sounds plausible, not what is true. Checking remains human work.

See also: Human in the loop, RAG (search, then answer)

A way of working where a person checks and approves what AI produces before anything happens. Sensible for anything that affects people, money or reputation.

Example: The AI drafts a reply to a complaint, an employee reads it and sends it.

See also: Hallucination, AI agent

Building software with blocks, forms and drag-and-drop instead of code. Low-code means a bit of code is added now and then. With AI, non-programmers can also build working tools.

See also: Vibe coding, Workflow and automation

The instruction or question you give an AI model. The clearer you say what you want, for whom, in what form and with what background, the better the result.

Example: Not 'write an email', but 'write a short, friendly email to a customer who pays late, with the invoice details below'.

See also: Prompt engineering, System prompt

The craft of writing and testing good prompts: giving context, adding examples, fixing the form of the answer and having the model work step by step. Increasingly part of everyday work rather than a separate job.

See also: Prompt

An approach where the AI first searches your own documents or a database, and only then answers based on what it found. That gives answers that match your information, with a reference to the source. RAG stands for retrieval-augmented generation.

Example: An internal assistant that answers staff questions from the employee handbook.

See also: Embedding, Hallucination

Fixed instructions that come before every conversation and that the user usually does not see. They set, for example, the role, the tone and what the model may and may not do.

Example: A customer service bot with the system prompt: 'You help customers of web shop X, in English, and you never give medical advice.'

Making software by describing in plain language what you want and letting AI write the code, without reading every line yourself. Fast for prototypes. For something that really has to keep working, checks, tests and security are still needed.

See also: No-code and low-code, Human in the loop

A series of steps that happen automatically one after another, often across different programs. With tools such as n8n, Make or Zapier you connect AI to your email, calendar, CRM or accounting, often without programming.

Example: Every new request through the contact form is summarised, labelled and added as a task to your planning.

See also: AI agent, API, No-code and low-code

A fixed way for programs to talk to each other. Through the API of OpenAI, Anthropic or Google you build AI into your own software or workflow, without the chat window.

See also: Workflow and automation, Inference

A text or image turned into a series of numbers that captures its meaning. Texts about the same thing get numbers that look alike. That lets software search by meaning instead of by the exact same words.

See also: RAG (search, then answer)

Training an existing model a little more with your own examples, so it gets better at one task or one style. A good prompt or RAG is often simpler and cheaper.

See also: Training, RAG (search, then answer)

Using a trained model: you ask something, the model computes an answer. With an API you usually pay for this per amount of tokens.

See also: API, Token

The way most AI learns: a computer finds patterns in many examples by itself, instead of someone writing down all the rules.

Example: A model that learns from thousands of past transactions which payments are suspicious.

See also: Training, Neural network and deep learning

An open standard that lets AI assistants work safely with other programs and data sources, such as your files, calendar or a database. A kind of universal plug between AI and your tools.

See also: AI agent, API

A neural network is a calculation model with many layers of connected 'nodes', loosely inspired by the brain. Deep learning is machine learning with such networks of many layers. Language models and image generators are built this way.

See also: Transformer

Models whose makers release the trained 'weights', so you can run them yourself, including on your own server. Llama, Mistral and Qwen are examples. Useful when data must not leave your own environment.

See also: Running locally

A language model that first 'thinks' in intermediate steps before answering. Slower and more expensive, but better at maths, planning, code and difficult trade-offs.

See also: Language model (LLM)

Using an AI model on your own computer or server instead of through an online service. More control over your data, but you need good hardware and the smaller models are less capable.

See also: Open models (open weights), GDPR and AI

A setting that controls how predictable or how varied a model's answers are. Low gives roughly the same answer every time, high gives more variety and creativity, but also more chance of nonsense.

The piece of text a language model computes with: a word, part of a word or a punctuation mark. In English, 100 words are roughly 130 tokens. Prices and limits of AI services are often expressed in tokens.

See also: Context window

The process in which a model learns from data. Training a large language model takes a lot of computing power, time and money and is done by the model makers. As a user you usually do not train the model yourself.

See also: Fine-tuning, Inference

The blueprint behind almost all modern language models, introduced in 2017. A transformer looks at every word in relation to all the other words in the text and weighs how much they matter to each other. The T in GPT stands for it.

See also: Language model (LLM)

The European law for AI. The higher the risk of an application, the stricter the rules. Some uses are banned, such as 'social scoring'. High-risk uses, such as AI in recruitment or in access to education, get strict requirements. The law comes into force step by step: the bans and the AI literacy duty apply since February 2025.

See also: AI literacy, High-risk AI

A skew in the results of AI, often because the training data was skewed. The model then takes over prejudices, for example about gender, origin or age. The Dutch childcare benefits scandal showed how severe the consequences of a skewed system can be.

See also: High-risk AI, Human in the loop

Much is still uncertain about copyright and AI. Two questions matter: were model makers allowed to use other people's work for training, and who owns what AI makes? Work made by AI without human creative choices is, according to many lawyers, not protected. In assignments, agree in writing who may do what.

See also: Generative AI

Images, audio or video made or edited with AI that look real, for example someone saying something they never said. Under the AI Act, such content must be recognisable as AI-made.

See also: Image generator, AI Act

Privacy law also applies when you use AI. If you put personal data into an AI service, you need to know where that data goes, whether it is used to train the model, and whether you have a good reason to do it. Business plans often offer more guarantees here than free versions.

See also: Running locally

Limits you build around an AI application: what the model may not say or do, which data it may not see, and when a person has to step in.

See also: System prompt, Human in the loop

AI that can have major consequences for people, for example when assessing job applicants, creditworthiness or access to care and benefits. Under the AI Act it must meet extra requirements for data, documentation, human oversight and transparency.

See also: AI Act, Bias