Plain-English definitions of the AI terminology behind the news — for an intelligent reader with no technical background.
An AI agent is a system that does not just answer questions but takes actions to get things done. Given a goal, it can plan a series of steps, use tools such as a web browser, a calendar, or another piece of software, check its own progress, and carry on until the task is complete. The shift is from an assistant that tells you how to do something to one that goes and does it for you.
AI alignment is the effort to make AI systems pursue what people actually intend, rather than something subtly or dangerously different. As models become more capable, the worry is not that they turn evil, but that they follow their instructions too literally or optimise for the wrong goal. Alignment research tries to ensure that an AI's behaviour reliably matches human values and intentions, and it is the foundation of most AI safety work.
When an AI system produces unfair or skewed results that disadvantage particular groups of people. It usually stems from patterns in the data a model learned from, and is one of the central ethical concerns in deploying AI.
A hypothetical AI that could match or exceed human ability across the full range of intellectual tasks, rather than being good at just one. AGI does not exist today, and experts disagree sharply on whether or when it might.
Shorthand for the raw computing power — chips, and the time spent running them — needed to train and run AI. Compute is one of the field's scarcest and most strategic resources, and the reason specialised chips and data centres matter so much.
Computer vision is the branch of AI that interprets images and video. It lets machines identify objects, read text in a photograph, detect faces, spot defects on a production line, or guide a self-driving car. Where a camera simply records light, computer vision works out what that light actually shows, turning raw pixels into useful understanding of a scene.
The amount of text an AI model can consider at one time — its short-term memory. Everything you send and everything it replies has to fit inside this window, measured in tokens. A larger context window lets a model work with longer documents and conversations.
Deep learning is a kind of machine learning that uses neural networks with many layers stacked on top of one another. The word "deep" refers to that depth, not to any special cleverness. Each layer learns to spot slightly more complex patterns than the one below it, so the network can build up from simple features to rich understanding. It is the approach behind most of the AI breakthroughs of the past decade, from image recognition to chatbots.
A diffusion model is a type of AI that generates images and video by learning to remove noise. During training it is shown pictures that have been gradually smothered with random visual static, and it learns to reverse that process step by step. To create something new, it starts from pure static and cleans it up until a coherent image emerges. This is the technology behind popular tools such as Stable Diffusion.
Fine-tuning means taking a model that has already been trained on a broad range of data and training it a bit further on a narrower, more specific set of examples. This specialises the model for a particular job, tone, or field without the enormous cost of building one from scratch. A general language model might be fine-tuned to sound like a legal assistant, to speak a company's house style, or to read medical notes.
A large, general-purpose AI model trained on a huge amount of data, built to be adapted to many different tasks rather than one. Think of it as a broadly educated graduate you can then train up for a specific job.
One of the largest, most capable AI models at the leading edge of what is currently possible. Frontier models are hugely expensive to build, made by only a handful of organisations, and the main focus of AI safety and regulation.
Generative AI describes systems that create brand new content rather than only analysing existing content. Give it a prompt and it can produce written text, images, music, speech, or video that did not exist before. It is the technology behind chatbots that write essays, tools that turn a sentence into a picture, and apps that clone a voice. The results are original in the sense that they are freshly assembled, though they are learned from huge amounts of existing human work.
The controls placed around an AI system to keep its behaviour within safe, acceptable limits — filtering harmful requests, blocking unsafe outputs, and keeping it on task. Guardrails sit around a model rather than inside it.
A way of building software that learns patterns from examples rather than following rules a programmer wrote by hand. Show it enough labelled photos of cats and it works out for itself what a cat tends to look like.
A model design that splits the network into many specialised sub-models, or "experts", and uses only the few most relevant ones for each input. This lets a model be very large in total while staying relatively cheap to run.
A technique for compressing a large, capable AI model into a smaller, faster one by training the small model to imitate the big one. Distillation is how much of a model's ability is packed into something cheap enough to run at scale or on a phone.
AI that can work with more than one kind of input or output — text, images, audio, and video — rather than just one. A multimodal model might describe a photo, answer a spoken question, or generate a picture from a written description.
Natural Language Processing, or NLP, is the branch of AI concerned with human language: reading it, understanding it, and producing it. It covers everything from spam filters and search engines to translation apps, voice assistants, and chatbots. The goal is to let computers work with the messy, ambiguous, endlessly varied way that people actually write and speak, rather than the rigid commands computers traditionally understand.
A neural network is a computer system loosely inspired by the way brain cells connect. It is built from layers of simple units, called nodes, that pass signals to one another. By adjusting the strength of those connections while learning from examples, the network gradually gets better at a task, whether that is recognising a face, translating a sentence, or predicting the next word. Almost all modern AI is built on neural networks of one kind or another.
Parameters are the adjustable internal values that a model learns during training. Each one is a small number that the model tunes as it works through examples, and together they store everything the model has learned. When you read that a system has "billions of parameters", that figure is a rough measure of its size and capacity, in the same spirit that engine size hints at the power of a car.
The craft of writing the instructions you give an AI model to get a better result. A prompt is simply what you type in; prompt engineering is the practice of phrasing, structuring, and giving enough context that the model does what you actually want.
A language model designed to work through a problem step by step before giving its answer, rather than replying instantly. This deliberate "thinking" makes it markedly better at maths, coding, and logic, at the cost of taking longer.
Deliberately attacking or probing an AI system to find its weaknesses before someone else does. Red teamers try to make a model misbehave — produce harmful content, leak data, or bypass its safeguards — so those flaws can be fixed.
A way of training AI by trial and error, rewarding good outcomes and penalising bad ones so the system gradually learns which actions work best. It is how AI masters games, controls robots, and, increasingly, learns to reason.
A training technique that uses people's judgements to teach a model which responses are helpful, honest, and appropriate. It is a large part of why modern chatbots feel polite and useful rather than merely fluent.
Retrieval-Augmented Generation, usually shortened to RAG, is a technique that lets an AI look things up before it answers. Instead of relying only on what it memorised during training, the system first searches a chosen knowledge base, such as a company's documents, for relevant information, then uses what it finds to write its reply. This makes answers more accurate, more current, and far less likely to be made up.
The technology that turns spoken language into written text. It powers voice assistants, live captions, meeting transcripts, and call-centre tools — and is a particular strength of several UK companies.
Artificially generated data used to train or test AI, rather than data collected from the real world. It helps when real data is scarce, sensitive, or expensive to gather — though it has to be made with care to be genuinely useful.
The small unit of text an AI language model actually reads and writes. Rather than whole words, models break text into tokens — often a word, part of a word, or a punctuation mark — and process one at a time. Token counts are how usage is measured and priced.
Training data is the collection of examples a model learns from. For a language model it might be billions of pages of text; for an image model, millions of pictures. The model studies this material to pick up patterns, so the quality, variety, and fairness of the data shape everything the model can do. As the old computing saying goes, garbage in, garbage out: a model is only ever as good as what it was taught with.
A transformer is the neural-network design that powers almost every modern language model, including the systems behind popular chatbots. Introduced by Google researchers in a 2017 paper titled "Attention Is All You Need", its key trick is called attention: the ability to weigh up how much each word in a sentence relates to every other word. This lets the model understand context and long-range meaning far better than earlier designs could.