AI glossary
Plain-language explanations of the AI words you meet in the news.
- AGI (artificial general intelligence)
- A hypothetical AI that can learn and do most intellectual tasks as well as a human. There is no agreed definition or test for it.
- AI agent
- An AI system that works towards a goal on its own: it plans steps, uses tools such as a browser or code, and checks the results.
- Alignment
- Work to make AI systems do what people intend and avoid harmful behaviour, even as they become more capable.
- API
- A way for programs to talk to each other. Developers use a company's API to build its AI models into their own apps, usually paying per token.
- Artificial intelligence (AI)
- Computer systems that do tasks we normally associate with human thinking, such as understanding language, recognising images or making decisions.
- Benchmark
- A standard test used to compare AI models. Results are useful, but a high score does not guarantee the model is best for your task.
- Chatbot
- A program you talk to in ordinary language. Modern AI chatbots use a large language model to write their answers.
- Context window
- How much text a model can take into account at once, counted in tokens. A larger window lets it read longer documents in one go.
- Data centre
- A building full of servers and chips. AI data centres need huge amounts of electricity and cooling.
- Deepfake
- A realistic fake video, image or voice made with AI, often of a real person. Deepfakes are used in scams and disinformation.
- Diffusion model
- A type of model that creates images, video or text by starting from noise and refining the result step by step.
- Fine-tuning
- Extra training of an existing model on a smaller, specific dataset so that it gets better at a particular task or style.
- Generative AI
- AI that creates new content – text, images, sound, video or code – instead of only sorting or analysing existing data.
- GPU (graphics processor)
- A graphics chip that can do many calculations in parallel. GPUs from companies like Nvidia are the main hardware for training and running AI.
- Hallucination
- When an AI model states something false with confidence, such as an invented fact or source. Always check important claims.
- Inference
- Running a trained model to get an answer. Every chat message is an inference, and it is what AI services pay for in chips and power.
- Large language model (LLM)
- An AI model trained on huge amounts of text to predict the next word. ChatGPT, Claude and Gemini are built on large language models.
- Mixture of experts (MoE)
- A model design that splits the network into many 'experts' and uses only a few for each word, making large models cheaper to run.
- Multimodal
- Able to handle several kinds of input or output, such as text, images, sound and video, in the same model.
- Open-weight model
- A model whose trained parameters are published so anyone can download and run it, for example on their own computer. Licences vary.
- Parameters
- The numbers inside a model that are adjusted during training. More parameters often mean more capability, but also more memory and cost.
- Prompt
- The instruction or question you give an AI model. Clear prompts with context and examples usually give better answers.
- Quantisation
- Storing a model's numbers with fewer bits so it needs less memory and runs faster, at the cost of a little quality. Common for running AI locally.
- RAG (retrieval-augmented generation)
- A technique where the AI first looks up relevant documents and then writes its answer from them, which reduces made-up facts.
- Reasoning model
- A model that works through a problem step by step before answering. It is often better at maths, logic and code, but slower and costlier.
- Token
- The small piece of text a language model reads and writes, often part of a word. Prices and limits are usually counted in tokens.
- Training
- The process where a model learns patterns from large amounts of data. Training a top model can take months and thousands of chips.