AI Glossary: Plain-English Explanations of Common AI Terms

The words that keep turning up in AI news, explained simply. Search for a term or jump to a letter.

90 terms · Definitions reflect the state of AI as of November 2026. Spotted something out of date? See our editorial standards.

A

Agent (AI agent)

An AI system that is given a goal and works towards it over several steps on its own, choosing which tools to use, such as web search, a browser or code, and checking its progress as it goes. A chatbot answers a message; an agent carries out a task.

Example: Asked to find and book a train, an agent might search timetables, compare fares, fill in the booking form and report back for your approval.

See also: Agentic AI, Tool use, Human in the loop

Agentic AI

A broad label for AI that acts rather than just responds: planning, taking actions in software and adapting to results with limited human input. It is often used loosely in marketing, so it is worth asking exactly what a product can do unsupervised.

Why it matters: The more an AI can act on its own, the more important permissions, spending limits and human sign-off become.

See also: Agent, Guardrails

AGI (artificial general intelligence)

A hypothetical AI that could match or exceed people across most intellectual tasks rather than in narrow areas. There is no agreed definition or test, and experts disagree sharply on whether and when it might arrive.

Why it matters: Several leading AI companies say building AGI is their goal, so the term shapes investment, regulation and safety debates.

See also: ASI, Frontier model

AI Act (EU)

The European Union's law regulating artificial intelligence, in force since August 2024. It sorts AI uses by risk: some are banned outright, "high-risk" uses such as recruitment or credit scoring face strict requirements, and makers of general-purpose models have transparency and safety duties. Its obligations are being phased in between 2025 and 2027, and the timing of some high-risk rules has been under review.

Why it matters: It applies to companies outside the EU, including UK firms, when they offer AI systems or their outputs in the EU. Fines for the most serious breaches can reach €35 million or 7% of global turnover.

See also: AI Security Institute, ICO

AI (artificial intelligence)

The broad field of building computer systems that do things which normally need human intelligence, such as understanding language, recognising images, making predictions or solving problems. Most AI in the news today is based on machine learning.

See also: Machine learning, Generative AI

AI safety

The field concerned with preventing AI from causing harm, from everyday problems like errors and misuse to larger worries about powerful systems acting against human interests. It covers testing, technical research and policy.

See also: Alignment, Red teaming, AI Security Institute

AI Security Institute (UK)

The UK government body that tests advanced AI models for risks to national security and public safety, such as help with cyber attacks or chemical and biological weapons. It was set up in 2023 as the AI Safety Institute after the Bletchley Park summit and renamed in February 2025. It sits within the Department for Science, Innovation and Technology.

Why it matters: It is the UK's main public source of technical expertise on frontier AI and works with AI companies to evaluate models before and after release.

See also: AI safety, Eval, Red teaming

AI slop

Slang for low-quality content churned out with AI, such as bland articles, fake-looking images or spammy videos, usually made in bulk to attract clicks.

Why it matters: Slop can crowd out reliable information in search results and social feeds.

See also: Generative AI, Model collapse

AI washing

Exaggerating or inventing the role of AI in a product or business to look more innovative. The term borrows from "greenwashing", and regulators in several countries have warned firms against misleading AI claims.

See also: Agentic AI

Alignment

The work of making AI systems pursue the goals their designers and users intend, and behave in line with human values, rather than finding unwanted shortcuts or causing harm. It becomes harder as systems grow more capable and autonomous.

See also: AI safety, RLHF

API (application programming interface)

A set of rules that lets one piece of software talk to another. AI companies sell access to their models through APIs, so other apps can send a request and get the model's answer back, usually paying per token.

Example: A customer service app might use an AI company's API behind the scenes to draft replies.

See also: SDK, Token

ASI (artificial superintelligence)

A hypothetical AI that would far surpass the best human minds in almost every field. It goes beyond AGI and remains speculative, though some AI leaders now openly say it is their aim.

See also: AGI, p

Attention

The technique inside a transformer that lets a model weigh how much each word in the input relates to every other word. It is what helps a model work out, for example, what "it" refers to in a long sentence.

See also: Transformer, Context window

B

Benchmark

A standard set of questions or tasks used to compare AI models, producing a score. Well-known examples test coding, maths, science knowledge or general reasoning.

Why it matters: Benchmark scores are useful but can mislead: models may have seen the test questions during training, and a high score does not guarantee good performance on your real tasks.

See also: Eval, SOTA

Bias

When an AI system produces unfair or skewed results for certain groups, often because its training data reflected existing inequalities or left some people out.

Example: A recruitment tool trained on past hiring decisions could learn to favour candidates who resemble those hired before.

See also: Training, Explainability

C

Chain of thought

Having a model work through a problem step by step in writing before giving its final answer. It tends to improve accuracy on maths, logic and planning, and is built into reasoning models.

Why it matters: The visible steps can help people check a model's work, although researchers caution that they do not always reflect how the model actually reached its answer.

See also: Reasoning model, Test-time compute

Chatbot

A program you talk to in everyday language through a chat window. Modern chatbots such as ChatGPT, Claude, Gemini and Copilot are built on large language models.

See also: LLM, Agent

Compute

Shorthand for the computing power, mainly specialised chips in data centres, used to train and run AI models. Access to compute is one of the biggest costs and bottlenecks in AI.

Why it matters: Governments, including the UK, now treat compute capacity as a matter of economic strategy.

See also: GPU, FLOPs, Data centre

Computer vision

The branch of AI that lets computers interpret images and video, for example recognising faces, reading number plates or spotting defects on a production line.

See also: Multimodal

Context engineering

The practice of carefully choosing what information, instructions and tools are fed into a model's context window so it can do a task well. The term gained ground in 2025 as a broader successor to "prompt engineering", especially for agents.

See also: Prompt, Context window, RAG

Context window

The maximum amount of text (measured in tokens) a model can consider at once, including your prompt, any documents and its own reply. Leading models in 2026 can handle hundreds of thousands of tokens, and some a million or more.

Why it matters: A bigger window lets a model work with long reports or whole codebases, but models can still miss details buried in very long inputs.

See also: Token, Context engineering

Copilot

An AI assistant that works alongside you inside other software, suggesting text, code or actions. It is also a brand name: Microsoft Copilot and GitHub Copilot are Microsoft's products.

See also: Chatbot, Agent

D

Data centre

A large building full of servers, storage and networking kit. AI data centres pack in thousands of GPUs and use huge amounts of electricity and, often, water for cooling.

Why it matters: Power and planning have become major hurdles. The UK government has designated "AI Growth Zones" to speed up new sites.

See also: Compute, Hyperscaler

Deep learning

A type of machine learning that uses neural networks with many layers. It powers almost all of today's leading AI, from speech recognition to chatbots.

See also: Neural network, Machine learning

Deepfake

Realistic fake video, images or audio made with AI to show someone saying or doing something they never did.

Why it matters: Deepfakes are used in fraud, harassment and disinformation. In the UK, sharing sexually explicit deepfakes of someone without consent is a criminal offence.

See also: Watermarking, Generative AI

Diffusion model

A kind of AI that creates images, video or audio by starting from random noise and gradually refining it into a picture that matches a text description. Many popular image and video generators use this approach.

See also: Generative AI, Multimodal

Distillation

Training a smaller, cheaper "student" model to imitate the answers of a larger "teacher" model, keeping much of its ability at a fraction of the running cost.

Why it matters: It is common practice within companies, but it has become contentious when firms accuse rivals of distilling their models without permission.

See also: Small language model, Quantisation

E

Edge AI

Running AI on or near the device where data is collected, such as a phone, camera, car or factory sensor, rather than in a distant data centre.

Why it matters: It can be faster, work without an internet connection and keep data more private.

See also: On-device AI, Small language model, Latency

Embeddings

Lists of numbers that represent the meaning of a piece of text, an image or another item, so that similar things end up with similar numbers. They let software search by meaning rather than exact keywords.

Example: A search for "cheap flights" can find a page about "budget airfares" because their embeddings are close.

See also: Vector database, RAG

Eval (evaluation)

A test designed to measure how an AI model performs or behaves on something specific, such as accuracy on a company's own tasks, or whether it will help with something dangerous. Benchmarks are one type of eval.

See also: Benchmark, Red teaming, AI Security Institute

Explainability

How far people can understand why an AI system produced a particular result. Large neural networks are often called "black boxes" because their reasoning is hard to inspect.

Why it matters: If AI helps decide on a loan, a job or a benefit claim, people need to understand and challenge the outcome. UK data protection law gives people rights around significant automated decisions about them.

See also: Bias, ICO

F

Few-shot and zero-shot

Zero-shot means asking a model to do a task with no examples; few-shot means including a handful of examples in the prompt so it can copy the pattern.

See also: Prompt

Fine-tuning

Taking a model that has already been trained and training it further on a smaller, specialised set of examples so it gets better at a particular task, style or subject.

Example: A law firm might fine-tune a model on its own templates so drafts follow its house style.

See also: Training, RLHF

FLOPs

Floating-point operations: the basic calculations computers perform. The total number used to train a model is a rough measure of how much compute went into it.

Why it matters: Regulators use it as a yardstick. Under the EU AI Act, general-purpose models trained with more than 1025 FLOPs are presumed to carry "systemic risk".

See also: Compute, AI Act

Foundation model

A large model trained on vast, broad data that can be adapted to many different tasks. Most chatbots and AI tools are built on top of one.

See also: Frontier model, LLM, Fine-tuning

Frontier model

One of the most capable AI models available at a given time, typically from labs such as OpenAI, Anthropic, Google DeepMind, Meta, xAI or leading Chinese developers.

Why it matters: Frontier models attract the closest safety scrutiny because new abilities, and new risks, tend to appear there first.

See also: Foundation model, AI Security Institute

G

Generative AI

AI that creates new content, such as text, images, music, video or code, in response to a request, rather than just sorting or analysing existing data.

See also: LLM, Diffusion model

GPU (graphics processing unit)

A chip originally designed for video game graphics that is very good at doing many calculations at once, which makes it ideal for AI. Nvidia dominates the market for AI GPUs.

See also: TPU, Compute

Grounding

Tying a model's answers to specific, reliable sources, such as search results or a company's documents, so it is less likely to make things up and can show where information came from.

See also: RAG, Hallucination

Guardrails

Rules and safety checks placed around an AI system to keep its behaviour within limits, for example blocking harmful requests, filtering outputs or requiring approval before an agent spends money.

See also: Human in the loop, Jailbreak, System prompt

H

Hallucination

When an AI model states something false or invented with apparent confidence, such as a fake quote, statistic or court case.

Why it matters: Hallucinations have become rarer in the best models but have not gone away, so important facts from AI should be checked.

See also: Grounding, RAG

Human in the loop

An approach in which a person must review or approve an AI's output or actions before they take effect, especially for serious steps like payments, deletions or messages to customers.

See also: Agent, Guardrails

Hyperscaler

One of the giant cloud computing companies, chiefly Amazon Web Services, Microsoft Azure and Google Cloud, that run vast networks of data centres and rent out computing power.

See also: Data centre, Compute

I

ICO (Information Commissioner's Office)

The UK's independent regulator for data protection and freedom of information. It enforces UK GDPR, including when personal data is used to train or run AI, and publishes guidance on AI and data protection. The Data (Use and Access) Act 2025 provides for it to be replaced by a board-led body, the Information Commission.

See also: Explainability, AI Act

Inference

Using a trained model to produce an answer, as opposed to training it. Every time you send a chatbot a message, it runs inference.

Why it matters: As AI use grows, the cost of inference, not just training, has become a big part of the industry's spending on chips and power.

See also: Training, Latency, Test-time compute

J

Jailbreak

A trick to get an AI model to ignore its safety rules, for example by role-play, unusual wording or long, confusing prompts.

See also: Guardrails, Prompt injection, Red teaming

L

Latency

The delay between sending a request and getting a response. Low latency matters for voice assistants, live translation and anything that needs to feel instant.

See also: Inference, Edge AI

LLM (large language model)

An AI model trained on enormous amounts of text to predict the next token, which lets it write, summarise, translate, answer questions and write code. GPT, Claude, Gemini and Llama are families of LLMs.

See also: Transformer, Token, Small language model

M

Machine learning

A way of building AI in which computers learn patterns from data rather than following rules written by hand. The more good examples a system sees, the better it usually gets.

See also: Deep learning, Training

MCP (Model Context Protocol)

An open standard, introduced by Anthropic in 2024 and since adopted widely across the industry, that gives AI assistants a common way to connect to outside tools and data, such as calendars, files or business software.

Why it matters: It works a bit like a universal plug, so a tool built once can be used by many different AI apps.

See also: Tool use, Agent, API

Mixture of experts (MoE)

A model design that splits the network into many specialised sub-networks ("experts") and only switches on a few of them for each piece of input. This makes very large models cheaper and faster to run.

See also: Parameters, Inference

Model

The trained system at the heart of an AI product: a large file of learned numbers (weights) plus the code that turns input into output. A single model can power many different apps.

See also: Weights, Parameters, LLM

Model collapse

A drop in quality that can happen when AI models are trained repeatedly on AI-generated content rather than human-made data, causing errors to build up and variety to shrink.

See also: Synthetic data, AI slop

Multimodal

Able to work with more than one type of input or output, such as text, images, audio and video. A multimodal assistant can, for example, describe a photo or hold a spoken conversation.

See also: Computer vision, LLM

N

Natural language processing (NLP)

The branch of AI concerned with understanding and producing human language, from translation and spam filters to today's chatbots.

See also: LLM

Neural network

A computing system loosely inspired by the brain, made of layers of connected nodes that adjust the strength of their connections as they learn from data.

See also: Deep learning, Parameters

O

On-device AI

AI that runs directly on your phone, laptop or other gadget instead of in the cloud. Phone makers increasingly build in smaller models and dedicated AI chips for this.

Why it matters: Your data does not have to leave the device, and features can work offline, though on-device models are usually less capable than the biggest cloud ones.

See also: Edge AI, Small language model

Open source AI

Strictly, AI released so that anyone can use, study, change and share it, with the code, the weights and enough information about the training data to rebuild it. The Open Source Initiative published a formal definition in 2024; relatively few major models meet it in full.

Why it matters: "Open source" is often used loosely for models that are really open weights.

See also: Open weights

Open weights

A model whose trained weights are published for anyone to download and run, often under licence conditions, but without necessarily sharing the training data or full code. Meta's Llama, Mistral's models, DeepSeek, Qwen and OpenAI's gpt-oss are examples.

Why it matters: Open-weights models can be run privately and customised, but once released they cannot be recalled, which is why they feature in safety debates.

See also: Open source AI, Weights

P

Parameters

The adjustable numbers inside a model that are tuned during training. Models are often described by their size in parameters, such as "70 billion parameters", though more is not always better.

See also: Weights, Model

p(doom)

Slang for someone's estimate of the probability that AI leads to catastrophe for humanity. It is a rough personal guess, not a scientific measurement.

See also: ASI, AI safety

Pre-training

The first and most expensive stage of building a model, in which it learns general patterns from a huge dataset. It is followed by "post-training", such as fine-tuning and RLHF, which shapes it into a helpful assistant.

See also: Training, Fine-tuning

Prompt

The instruction or question you give an AI model. Clear prompts with context and examples usually get better results; crafting them carefully is known as prompt engineering.

See also: System prompt, Context engineering, Few-shot and zero-shot

Prompt injection

An attack in which instructions are hidden in content an AI reads, such as a web page, email or document, in the hope it will follow them instead of the user's instructions.

Why it matters: It is one of the main security risks for AI agents that browse the web or handle email, and there is no complete fix yet.

See also: Jailbreak, Agent, Guardrails

Q

Quantisation

Shrinking a model by storing its numbers with less precision, for example 4 or 8 bits instead of 16. It makes models smaller and faster, with some loss of quality.

See also: Distillation, On-device AI

R

RAG (retrieval-augmented generation)

A technique in which an AI system first looks up relevant information, for example from a company's documents or the web, and then uses it to write its answer.

Why it matters: It keeps answers up to date and grounded in real sources without retraining the model, and lets it cite where information came from.

See also: Embeddings, Vector database, Grounding

Reasoning model

A model trained to spend extra time working through a problem step by step before answering. Since OpenAI's o1 in 2024, most major labs have released them.

Why it matters: They do better on maths, coding and complex planning, but are slower and can cost more per answer.

See also: Chain of thought, Test-time compute

Red teaming

Deliberately attacking or probing an AI system to find weaknesses, harmful behaviours or security holes before bad actors do.

See also: Eval, Jailbreak, AI safety

RLHF (reinforcement learning from human feedback)

A training method in which people rate or rank a model's answers, and the model is adjusted to produce more of the kinds of answers people prefer. It helped make chatbots such as ChatGPT more helpful and polite.

See also: Pre-training, Alignment, Sycophancy

S

Scaling laws

The observed pattern that AI models tend to improve predictably as they are given more data, more parameters and more compute. Whether this will continue, and at what cost, is one of the industry's big debates.

See also: Compute, Test-time compute

SDK (software development kit)

A package of code libraries, tools and documentation that makes it easier for developers to build on a platform, such as an AI company's API.

See also: API

Small language model (SLM)

A compact language model, typically with a few billion parameters or fewer, designed to be cheap and fast enough to run on a laptop, phone or modest server. Microsoft's Phi and Google's Gemma are examples.

See also: LLM, On-device AI, Distillation

SOTA (state of the art)

Jargon for the best result achieved so far, as in "a new SOTA on this benchmark".

See also: Benchmark

Sovereign AI

A country's effort to have its own AI capability, such as domestic data centres, compute, models or expertise, so it is less dependent on foreign companies or governments.

See also: Compute, Data centre

Stochastic parrot

A critical phrase, from a 2021 research paper, for the idea that language models stitch together plausible-sounding text based on statistics, without real understanding. How well it describes today's models is hotly debated.

See also: LLM

Sycophancy

When an AI model tells users what they want to hear, such as agreeing with mistaken claims or flattering them, instead of giving an honest answer.

Why it matters: It can reinforce bad decisions or false beliefs. OpenAI rolled back a ChatGPT update in 2025 after complaints that it had become too flattering.

See also: RLHF, Alignment

Synthetic data

Data created by computers, often by other AI models, rather than collected from the real world. It is used to train models when real data is scarce, expensive or sensitive.

See also: Training, Model collapse

System prompt

Hidden background instructions given to a model by the company or developer, setting its role, tone and rules before you type anything.

See also: Prompt, Guardrails

T

Temperature

A setting that controls how random a model's output is. Low temperature gives more predictable, consistent answers; high temperature gives more varied, creative ones.

See also: Inference

Test-time compute

Extra computing power used while a model is answering, letting it think longer, try several approaches or check its work. It is the main idea behind reasoning models.

See also: Reasoning model, Inference, Scaling laws

Token

The chunk of text a language model reads and writes, often part of a word. In English, 1,000 tokens is roughly 750 words. AI services usually price usage per million tokens.

See also: Context window, API

Tool use

A model's ability to call on outside software, such as a search engine, calculator, code runner or booking system, to get information or take actions it could not manage alone.

See also: Agent, MCP

TPU (tensor processing unit)

A chip designed by Google specifically for AI work. Google uses TPUs to train and run its own models and rents them to other companies through Google Cloud.

See also: GPU, Compute

Training

The process of teaching a model by showing it vast amounts of data and repeatedly adjusting its parameters to reduce its mistakes. Training a frontier model can take months and cost hundreds of millions of pounds.

See also: Pre-training, Fine-tuning, Inference

Transformer

The neural network design behind almost all modern language models, introduced by Google researchers in 2017. Its key ingredient is attention. The "T" in GPT stands for transformer.

See also: Attention, LLM

V

Vector database

A database built to store embeddings and quickly find the ones most similar in meaning to a query. It is a common building block in RAG systems.

See also: Embeddings, RAG

Vibe coding

Building software by describing what you want to an AI in plain language and accepting the code it writes, often without reading it closely. The phrase was coined by AI researcher Andrej Karpathy in February 2025 and was Collins Dictionary's Word of the Year for 2025.

Why it matters: It lets non-programmers build apps quickly, but code nobody has checked can hide bugs and security holes.

See also: Agent, Copilot

W

Watermarking

Embedding a hidden signal or attached label in AI-generated images, audio, video or text so it can later be identified as AI-made. Examples include Google DeepMind's SynthID and the Content Credentials standard.

Why it matters: Watermarks can help tackle deepfakes, but many can be removed or are not used by every AI tool.

See also: Deepfake, Generative AI

Weights

The learned numbers that make up a trained model; in practice the word is often used interchangeably with parameters. Whoever has the weights can run the model.

See also: Parameters, Open weights

World model

An AI system that builds an internal picture of how the physical world works, so it can predict what happens next or generate interactive, explorable environments. World models are a growing research area for robotics and video generation.

See also: Multimodal