AI jargon, in plain English

A to Z

A

Agent

An AI tool that can carry out a series of steps on its own to finish a task, such as searching the web, filling in a form and drafting a summary, rather than just answering one question. Agents can save time, but they can also make mistakes at each step, so anything important still needs a person to check it.

AI (artificial intelligence)

Computer systems that do things we'd normally think need human intelligence: understanding language, recognising pictures, spotting patterns or making predictions. Most of the AI people talk about today is generative AI, which creates text, images and sound.

AI assistant

A general name for tools you can ask questions or give jobs to in everyday language, such as ChatGPT, Gemini, Copilot, Claude, Siri and Alexa. Some live in an app, some are built into your phone, email or office software.

AI literacy

Knowing enough about AI to use it sensibly: what it's good at, where it goes wrong, and what information you shouldn't put into it. The EU's AI Act asks organisations that use AI to take steps to support their staff's AI literacy, which can include UK firms whose AI use reaches customers in the EU.

AI policy

A short set of rules for how a business or organisation uses AI: which tools are allowed, what must never be typed in, what must be checked by a person, and who to ask. See our page on getting an AI policy for your business.

Algorithm

A set of step-by-step instructions a computer follows to solve a problem. A recipe is an algorithm for a cake. In AI, the word is often used loosely for the system that decides what you see, such as the posts in a social media feed.

API

A way for one piece of software to talk to another. When a booking system sends a customer's question to an AI model and gets an answer back, it usually does it through an API, short for application programming interface.

Automation

Setting software to do a routine job by itself, such as copying the details from every web enquiry into a spreadsheet. Adding AI to an automation lets it handle messier work, like drafting a reply to each enquiry for a person to approve.

B

Bias

When an AI system treats some people or groups unfairly, usually because the data it learned from was unbalanced. For example, a tool trained mostly on one kind of CV might rate other applicants lower. It's one reason important decisions shouldn't be left to AI alone.

C

Chatbot

A program you have a typed or spoken conversation with. Older chatbots followed fixed scripts; modern ones use large language models, so they can handle questions they weren't specifically prepared for, and can also get things confidently wrong.

ChatGPT

The AI assistant made by OpenAI, launched in November 2022. It's free to use with limits, with paid plans for heavier use and for businesses. The name comes from GPT, the family of AI models behind it.

Claude

An AI assistant made by the company Anthropic, available free with limits, with paid plans for individuals and teams. It's popular for writing, working with long documents and computer code.

Context window

How much text an AI model can take into account at once, including your question, any documents you've given it and the conversation so far. Very long chats can push early details out of view, which is why the AI sometimes forgets what you said at the start.

Copilot

Microsoft's AI assistant. There's a free version for anyone, a version inside Microsoft 365 business accounts, and a paid version that works inside Word, Excel, Outlook and Teams with your own files. The name is also used loosely for any AI helper that works alongside you.

D

Data protection

The law on how personal information is collected, stored and used, which in the UK means UK GDPR and the Data Protection Act. It applies to AI tools too: putting customers' details into a public AI tool can break it.

Deep learning

A type of machine learning that uses many layers of artificial neurons to learn complicated patterns, such as how faces or voices vary. It's the technique behind most of today's AI.

Deepfake

A fake video, photo or voice recording made with AI to look or sound like a real person. Criminals use deepfakes in scams, for example a video call where the 'boss' asks for an urgent payment. A family safe word or a call back on a known number is a simple defence.

F

Fine-tuning

Training an existing AI model a little further on a particular set of examples, so it gets better at one job or one style. Most small businesses never need it: giving the AI your documents and clear instructions usually does the job.

G

Gem

A custom version of Gemini that you set up with your own instructions and files, so it does one job well, such as writing in your business's style.

Gemini

Google's AI assistant and the name of its AI models. It's available as a free app, built into many Android phones, and inside Gmail, Docs and other Google Workspace apps.

Gemini Notebook

A free Google tool, called NotebookLM until July 2026, that answers questions using only the documents you give it, such as a manual, a report or your notes, and shows which part of each document each answer came from.

Generative AI

AI that creates new content, such as text, images, music, voices or video, in response to a request. ChatGPT writing a letter and an app turning a sketch into a picture are both generative AI.

GPT

Short for generative pre-trained transformer: OpenAI's family of AI models that power ChatGPT. It's often used, wrongly, as a general word for any AI chatbot.

Guardrails

Rules built into an AI tool to stop it doing harmful things, such as refusing to help write malware. They aren't perfect, so you still need your own common sense and checks.

H

Hallucination

When an AI states something false as if it were fact: a made-up statistic, a book that doesn't exist, or a law that was never passed. It happens because the AI predicts likely-sounding words rather than looking facts up. Always check anything that matters.

I

Image generator

An AI tool that creates pictures from a written description, such as 'a watercolour of Whitstable harbour at sunset'. Most AI assistants can now do this, usually with a daily limit on free plans.

L

Large language model (LLM)

The kind of AI model behind ChatGPT, Gemini, Claude and Copilot. It has learned patterns from huge amounts of text and uses them to predict what words should come next, which lets it write, summarise, translate and answer questions.

M

Machine learning

A way of building AI where the computer learns patterns from examples instead of following rules a person wrote. Spam filters learn what junk email looks like from millions of examples.

Memory

A feature in some AI assistants that remembers details from earlier conversations, such as your name or your business, so you don't have to repeat them. You can usually see, edit or switch it off in the settings.

Model

The trained AI system itself, the 'brain' behind an app. The same model can sit behind several apps, and one app can offer a choice of models, some faster and some more careful.

Multimodal

Able to work with more than one kind of input or output: text, pictures, sound and video. A multimodal assistant can look at a photo of your boiler's error light and talk you through what it means.

N

Natural language

Ordinary human language, the way you'd say or write something to a person. The big change with modern AI is that you can give instructions in natural language rather than learning special commands.

Neural network

A computer system loosely inspired by the brain, made of many connected units that adjust as they learn from examples. Large language models are very large neural networks.

O

Open-source model

An AI model whose workings are published so anyone can download, run and adapt it, sometimes on their own computer. It can help with privacy, but it takes more technical work to set up safely.

P

Phishing

Messages that pretend to be from a trusted company or person to trick you into clicking a link, paying money or giving away details. AI makes phishing messages more convincing, because the spelling and grammar are now usually perfect.

Plugin

An add-on that connects an AI assistant to another service or gives it a new skill, such as reading your calendar or creating a spreadsheet. Only connect services you trust, and check what each plugin can see.

Project

In ChatGPT and Claude, a workspace where you keep related chats, files and instructions together, so the AI has the background it needs each time, such as your price list and house style.

Prompt

What you type or say to an AI tool to tell it what you want. Clear prompts get better results: say who it's for, what you want, how long, and in what style.

Prompt injection

A trick where hidden instructions inside a web page, email or document try to make an AI tool do something it shouldn't, such as reveal information. It's one reason to be careful about letting AI agents act on your accounts.

R

RAG (retrieval-augmented generation)

A technique where an AI looks up information in a set of documents before answering, so the answer is based on your own material, for example a website assistant that answers from your FAQ and price list.

Reasoning model

An AI model that works through a problem step by step before answering, which takes longer but tends to be more reliable on maths, planning and tricky questions. Some apps show this as a 'think' or 'thinking' option.

S

Scam detection

Features on phones and in email that use AI to spot likely scam calls and messages and warn you. They help, but they don't catch everything.

Synthetic media

Images, video, audio or text created or changed by AI. Deepfakes are one kind; an AI-edited holiday photo is another. Some platforms now label it.

T

Token

The small chunks of text an AI model reads and writes, roughly three-quarters of a word each in English. Usage limits and prices for AI services are often counted in tokens.

Training

The process of building an AI model by showing it vast numbers of examples so it learns patterns. When a company says your chats 'may be used for training', it means they might help improve future models. Business plans usually switch this off.

Training data

The text, images and other material an AI model learned from. Its quality and balance shape what the AI knows and what it gets wrong.

Transcription

Turning speech into written text. AI does this quickly for meetings, calls and voice notes. Tell people before you record or transcribe a conversation.

Turing test

An old idea for judging whether a machine can hold a conversation well enough to pass for a human. Modern chatbots often can, which is exactly why AI scams work.

V

Voice assistant

An AI you talk to out loud, such as Siri, Alexa, Google Assistant or the voice mode in ChatGPT and Gemini. Newer ones can hold a natural back-and-forth conversation.

Voice cloning

Using AI to copy a person's voice from a short recording, so it can say anything. It has good uses, such as helping people who are losing their voice, and criminal ones, such as fake calls from a 'relative' in trouble.

W

Watermark

A hidden or visible mark added to AI-generated content to show where it came from. Some AI companies add invisible watermarks to images and audio, but they can't be relied on to catch every fake.

Workflow

The series of steps a job goes through, such as enquiry, quote, booking, invoice. Mapping a workflow is usually the first step in deciding where AI or automation could help.

Z

Zero-shot and few-shot

Asking an AI to do something with no examples (zero-shot) or with a few examples to copy (few-shot). Showing two or three examples of what you want is one of the easiest ways to get better results.