AI words in plain English

The ten you'll hear most

If you only learn ten, learn these. The rest of the page is in alphabetical order.

AI (artificial intelligence)

Computer programs that do things we'd normally say need thinking: writing, recognising pictures, answering questions.The thing that suggests the next word when you text is a small AI.

Chatbot

A program you talk to by typing or speaking, which replies in sentences.ChatGPT, Claude, Gemini and Copilot are all chatbots.

Prompt

The message you type to an AI tool. That's all. A better prompt gets a better answer."Write a polite reply saying we're closed on Mondays" is a prompt.

Model

The "brain" behind an AI tool: a very large program that has learned patterns from huge amounts of text, pictures or sound.ChatGPT is the app. The model is the engine inside it.

LLM (large language model)

A model trained on enormous amounts of text, so it can read and write language. It's the kind of model behind chatbots.When someone says "an LLM", they mean the engine inside tools like ChatGPT or Claude.

Generative AI

AI that makes new things, such as text, pictures, sound or video, rather than only sorting or labelling what already exists.Asking for a poem, or a picture of a cat in a hat.

Hallucination

When an AI tool states something false as if it were true. It isn't lying. It writes what sounds likely, and sometimes that's wrong.It gives you a book title and author that don't exist.

Training

How a model learns: it's shown huge numbers of examples and adjusts itself until it gets good at predicting what comes next.Like learning a language by reading millions of pages, without ever being taught the grammar.

Agent

An AI tool that doesn't only answer but takes steps for you: searching, clicking, filling in forms, then checking its work."Find three hotels under £100 near the station and put them in a table" done for you, start to finish.

Token

The small pieces of text a model reads and writes: roughly a short word or part of a word. Limits and prices are counted in tokens."Unbelievable" might be three tokens: un, believ, able.

A to Z

Algorithm

A set of step-by-step instructions a computer follows. Not all algorithms are AI.A recipe is an algorithm for a cake.

API

A way for one program to talk to another. It's how other apps plug an AI model into their own product.Your accounting app using an AI model behind the scenes to read receipts.

Automation

Getting a computer to do a repeated job without you. It may or may not use AI.An email that sends itself every time someone fills in your form.

Bias

When an AI tool treats people or ideas unfairly because of patterns in what it learned from.A tool that assumes every nurse is a woman.

Context window

How much text a model can keep in mind at once: your messages, its replies and anything you've pasted. Go past it and it starts to forget the earliest parts.In a very long chat, it loses track of what you said at the start.

Copilot

Microsoft's name for its AI assistants, built into Word, Excel, Outlook and Windows. People also use the word for any AI helper that works alongside you.Asking Word to draft a letter from your notes.

Data

Information a computer can use: words, numbers, pictures, clicks.Your sales spreadsheet is data. So is a folder of photos.

Deepfake

A fake video, picture or voice recording made with AI to look or sound like a real person.A video of a celebrity saying something they never said.

Fine-tuning

Giving a ready-made model extra training on your own examples so it's better at one particular job.Teaching a general model to write in your company's style.

GPT

A family of models made by OpenAI. It stands for "generative pre-trained transformer". ChatGPT is the app built on them."GPT" is the engine. "ChatGPT" is the car.

Guardrails

Rules built into an AI tool to stop it doing harmful things or going off-topic.It refuses to help write a threatening message.

Human in the loop

A person checks or approves what the AI does before it counts.The AI drafts the email. You read it and press send.

Image generator

An AI tool that makes pictures from a written description.Typing "a watercolour of a lighthouse at sunset" and getting a picture.

Inference

The moment a trained model is actually used: you ask, it works out an answer.Every time you press Enter in a chatbot, that's inference.

Knowledge cutoff

The date a model's training stopped. It may not know about anything after that unless it can search the web.Asking about last week's news and getting an out-of-date answer.

Machine learning

The broader method behind modern AI: programs that learn from examples and aren't given every rule by hand.A spam filter that gets better as you mark emails as junk.

Memory

A feature in some chatbots that remembers things about you between chats, such as your job or how you like answers written. You can usually see and delete what it has saved.It remembers you're vegetarian when you ask for recipes next month.

Multimodal

An AI tool that works with more than one kind of thing: text, pictures, sound, video.Taking a photo of a letter and asking what it means.

Neural network

The design most modern AI is built on: layers of simple maths loosely inspired by how brain cells connect.You never need to see one, just as you never see the wiring behind a light switch.

Open source

Software, or a model, whose inner workings are published so anyone can inspect, use or change them.A recipe that's shared freely, compared with a secret family one.

Parameters

The millions or billions of internal dials a model adjusts during training. More isn't always better."A 70-billion-parameter model" is a way of saying it's large.

Personal data

Any information about a living person who could be identified from it. The law limits what you can do with it.A customer's name and address. A pupil's grades.

Prompt engineering

A grand name for writing clear instructions for an AI tool.Saying who it's for, what you want and how long it should be.

RAG (retrieval-augmented generation)

A way of making an AI tool look things up in a set of documents before it answers, so the answer is based on them.A company chatbot that answers from the staff handbook.

Reasoning model

A model that works through a problem in steps before answering. Slower, and often better at tricky questions.Planning a week of meals for four people with different diets on a budget.

Speech-to-text

Turning spoken words into written text. Also called transcription or dictation.Talking to your phone and seeing the words appear.

System prompt

Hidden instructions the maker of an AI tool gives it before you type anything, setting how it should behave.Why a shop's chatbot only talks about the shop.

Temperature

A setting for how predictable or varied a model's answers are. Most people never need to touch it.Low for a contract summary. Higher for brainstorming names.

Text-to-speech

Turning written text into a spoken voice.Your phone reading a message aloud while you drive.

Training data

The examples a model learned from.Books, websites and articles, for a language model.

Transformer

The design behind today's language models, introduced by researchers at Google in 2017. It's the "T" in GPT.You'll hear the word. You won't need to use one.

Voice mode

Talking to a chatbot out loud and hearing it reply, like a phone call.Practising a job interview on your walk to the station.

Workflow

The steps a job goes through from start to finish.Getting a booking, confirming it, sending a reminder, asking for a review.