The AI jargon buster.

A

Agent

An AI system that can take actions to complete a task, not just answer a question. For example, it might search, fill in a form or update a record, working through several steps on its own. Read more →

Agentic AI

The general term for AI that plans and carries out multi-step tasks using tools and other systems. It can save a lot of effort, but it needs clear limits and a person checking what it does. Read more →

AI (artificial intelligence)

A broad label for computer systems that do things we normally associate with human intelligence, such as understanding language, recognising images, spotting patterns or making predictions.

AI literacy

A working understanding of what AI is, what it can and can't do, and how to use it safely and sensibly. It's about good judgement, not technical skill.

Algorithm

A set of step-by-step instructions a computer follows to solve a problem or make a decision. AI models are built and trained using algorithms.

API (application programming interface)

A way for one piece of software to talk to another. Many AI tools are reached through an API, which is how they get built into other apps and systems.

B

Bias

When an AI system treats some people or groups unfairly, usually because the data it learned from reflected unfair patterns. It matters most in decisions about people, such as hiring or lending. Read more →

C

Chatbot

A program you talk to by typing or speaking, which replies in ordinary language. Modern chatbots such as ChatGPT, Copilot, Gemini and Claude are built on large language models. Read more →

Computer vision

AI that works with images and video: reading text from scanned documents, recognising objects, checking for damage or counting items. Read more →

Context window

How much text an AI model can take in and keep track of at once, including your conversation, any documents you add and its own replies. Anything beyond that limit drops out of view.

Copilot

A general name for an AI assistant built into software you already use, such as email, documents or spreadsheets. It's also the brand name Microsoft uses for its AI assistants. Read more →

D

Data lake

One big store where an organisation keeps lots of its data, in whatever shape it comes: spreadsheets, documents, emails, logs. AI tools can draw on it, but only if someone owns it and keeps it tidy, or it just moves the mess. Read more →

Data protection / GDPR

The rules on how personal information is collected and used. In the UK these are mainly UK GDPR and the Data Protection Act 2018, and they apply to AI tools just as they do to any other system. Read more →

Data warehouse

A store of an organisation's data that has already been cleaned and organised into neat tables, ready for reports and analysis. Tidier than a data lake, but less flexible. Read more →

Deep learning

A type of machine learning that uses large neural networks with many layers. It's the approach behind most modern AI, including chatbots, image recognition and speech recognition. Read more →

Deepfake

A fake but realistic image, video or audio clip made with AI, often imitating a real person. Deepfakes are increasingly used in scams, such as a cloned voice asking for an urgent payment. Read more →

E

Embedding

A way of turning text or images into a list of numbers that captures their meaning, so that similar things end up close together. It lets search tools find matches by meaning rather than exact words. Read more →

EU AI Act

A European Union law that regulates AI according to risk, with stricter rules for high-risk uses such as recruitment. Its requirements are coming in over several years, and it can apply to UK organisations that do business in the EU. Read more →

Explainability

How well people can understand why an AI system produced a particular answer or decision. It matters when you need to justify a decision to a customer, a regulator or a colleague. Read more →

F

Fine-tuning

Giving an existing AI model extra training on a smaller, specialised set of examples, so it gets better at a particular task, style or subject. Read more →

Foundation model

A large, general-purpose AI model trained on huge amounts of data, which can then be adapted for many different tasks. The models behind today's chatbots are foundation models. Read more →

G

Generative AI

AI that creates new content, such as text, images, audio, video or computer code, in response to a request. Chatbots are the best-known example. Read more →

GPU (graphics processing unit)

A computer chip originally designed for graphics, which is very good at doing many calculations at once. GPUs are the main hardware used to train and run modern AI. Read more →

Grounding

Making an AI tool base its answers on specific, trusted sources, such as your organisation's documents, rather than only on what it learned in training. It reduces, but doesn't remove, made-up answers. Read more →

Guardrails

Rules and checks built around an AI system to stop it producing harmful, off-topic or unsafe output, or taking actions it shouldn't. Read more →

H

Hallucination

When an AI tool states something false as if it were true, such as an invented fact, quote, figure or reference. It happens with every tool, which is why checking matters. Read more →

Human in the loop

Designing a process so that a person reviews or approves what AI does before it takes effect, especially for important or risky decisions. Read more →

I

Inference

Using a trained AI model to produce an answer. Every time you ask a chatbot something, that's inference. It's separate from training, which is how the model learned in the first place. Read more →

Intelligent automation

Combining AI with traditional automation to handle routine work, such as reading invoices, sorting emails or updating records, with people stepping in for the exceptions. Read more →

K

Knowledge cutoff

The date after which an AI model has no training information. It won't know about later events unless the tool can search the web or look at documents you give it.

L

LLM (large language model)

An AI model trained on an enormous amount of text to predict and produce language. LLMs power chatbots and most AI writing tools. Read more →

M

Machine learning

The main way modern AI is built: instead of being given exact rules, a computer learns patterns from examples and uses them to make predictions or decisions. Read more →

MCP (Model Context Protocol)

An open standard, first released by Anthropic in 2024 and now widely used, for connecting AI assistants to other tools and data sources. Think of it as a universal plug. Read more →

Model

The trained program at the heart of an AI tool: the thing that has learned patterns from data and uses them to produce answers. One model can power many different apps.

Multimodal

Able to work with more than one kind of information, such as text, images, audio and video. A multimodal tool could describe a photo or answer a question about a chart. Read more →

N

Natural language processing (NLP)

The branch of AI that deals with human language: understanding it, sorting it, translating it and producing it. Read more →

Neural network

A kind of AI model loosely inspired by the brain, made of many connected layers of simple calculations. It learns by adjusting the strength of those connections during training. Read more →

O

Open-source model

An AI model whose inner workings (its weights) are published so anyone can download, run and adapt it. Many are more accurately called “open-weight”, because the training data isn't shared.

P

Parameters

The internal numbers a model adjusts as it learns, which together hold what it has learned. Large models have billions of them. More isn't always better. Read more →

Predictive analytics

Using data and statistics, often with machine learning, to forecast what's likely to happen, such as demand, staffing needs or late payments. Read more →

Prompt

The question or instruction you give an AI tool. Clear prompts with context about who it's for and what good looks like get much better results. Read more →

Prompt injection

A trick where instructions are hidden in an email, document or web page so that an AI tool reading it does something it shouldn't, such as leaking information. It's a growing risk as AI tools get more access. Read more →

R

RAG (retrieval-augmented generation)

A way of making an AI tool look up relevant information, such as your organisation's documents, before it answers, so its replies are based on those sources. Many “chat with your documents” tools use it. Read more →

Reasoning model

An AI model designed to work through a problem step by step before answering. It's usually slower but better at maths, logic and planning, though it can still be wrong. Read more →

Reinforcement learning

A way of training AI by trial and error, rewarding good results and discouraging bad ones. It's also used, with feedback from people, to make chatbots more helpful and safer. Read more →

Responsible AI

Using AI in a way that is fair, safe, transparent and accountable, with clear human responsibility for the outcomes. Read more →

S

Shadow AI

Staff using AI tools at work that their organisation hasn't approved, often with good intentions. It's a common way for confidential data to leak. Read more →

Supervised learning

Machine learning from labelled examples, where each example comes with the right answer. For instance, emails marked “spam” or “not spam” teach a spam filter. Read more →

System prompt

Background instructions set by whoever built an AI tool, which shape how it behaves before you type anything, such as its tone, its rules and what it should refuse to do. Read more →

T

Temperature

A setting that controls how predictable or varied an AI model's output is. Low temperature gives steadier, more repeatable answers. High temperature gives more variety. Read more →

Token

A small chunk of text, often a word or part of a word, that an AI model reads and writes. Usage limits and costs for AI tools are usually measured in tokens. Read more →

Training data

The examples an AI model learns from. Its quality, age and gaps shape what the model can do, and what mistakes or biases it picks up. Read more →

U

Unsupervised learning

Machine learning that finds patterns in data without being given the answers, such as grouping customers with similar buying habits. Read more →

V

Vector database

A database that stores embeddings, so it can quickly find the items closest in meaning to a question. It's often used to help AI tools search an organisation's documents. Read more →