AI glossary
Short explanations of words and phrases you may encounter when reading or hearing about artificial intelligence.
21 terms, in alphabetical order
- Agent
An AI system that can work through several steps towards a goal, often using tools and checking what happened.
Example: An agent could compare train times, check your calendar and prepare an itinerary for you to approve.
Related: Tool, Model
- Artificial intelligence (AI)
A broad name for computer systems designed to do tasks that usually need human abilities, such as recognising speech or finding patterns.
Example: A phone that turns spoken words into captions is using one type of AI.
Related: Machine learning, Model
- Generative AI
AI that creates new material, such as text, pictures, audio or video, based on patterns in its training data.
Example: Asking a chatbot to draft a birthday invitation uses generative AI.
Related: Large language model, Training data
- Graphics processing unit (GPU)
A computer processor designed to perform many calculations in parallel. GPUs were first developed for graphics and are now widely used to train and run AI models.
Example: Training a large language model may use thousands of powerful GPUs working together.
Related: Model, Pretraining
- Hallucination
When a generative AI system gives an answer that sounds plausible but is incorrect or unsupported.
Example: A chatbot may confidently give the title of a book that does not exist.
Related: Generative AI, Prompt
- Large language model (LLM)
A model trained on a very large amount of text so it can work with language by predicting likely sequences of words or word-parts.
Example: Large language models are used in assistants such as ChatGPT, Claude and Gemini.
Related: Model, Token, Training data
- Machine learning
A way of building software by helping a computer find patterns in examples instead of writing every rule by hand.
Example: A spam filter can learn patterns from messages previously marked as spam.
Related: Artificial intelligence (AI), Training data
- Model
A trained computer system that uses patterns to produce an output from an input, such as labelling a photo or answering a question.
Example: A weather model uses observations as inputs and produces a forecast as its output.
Related: Machine learning, Training data
- Model Context Protocol (MCP)
An open standard that lets AI applications connect to external information, tools and services through a common interface.
Example: An AI assistant could use an MCP connection to read a calendar or search an organisation’s approved documents, with the appropriate permission.
Related: Agent, Tool, Model
- Multimodal
Able to work with more than one type of information, such as text, images and sound.
Example: A multimodal assistant might answer a written question about a photograph you provide.
Related: Model, Generative AI
- Open-source AI
AI made available on terms that allow people to use, study, change and share it, with enough access to understand how the system was created.
Example: A research group might publish a model, its code and detailed training information so others can examine and adapt the system.
Related: Model, Open-weight model, Training data
- Open-weight model
A model whose learned numerical parameters, called weights, can be downloaded. Its code, training data or licence may still have restrictions.
Example: A company may allow people to run a model on their own computers without publishing the data used to train it.
Related: Model, Open-source AI, Training data
- Post-training
Training and adjustment carried out after pretraining to make a model more useful, specialised or better aligned with instructions.
Example: A general language model may go through post-training so it becomes better at following questions in a chat.
Related: Pretraining, Reinforcement learning from human feedback (RLHF), Model
- Pretraining
The initial stage in which a model learns broad patterns from a large collection of examples before being adapted for particular tasks.
Example: During pretraining, a language model learns patterns from large amounts of text by repeatedly predicting missing or next tokens.
Related: Post-training, Training data, Token
- Prompt
The instruction or information a person gives to a generative AI system.
Example: “Explain this letter in plain English” is a prompt.
Related: Generative AI, Hallucination
- Reinforcement learning
A way of training a system through feedback about the results of its actions. Results that receive a higher reward become more likely.
Example: A game-playing system can receive a reward for winning and use repeated attempts to learn which actions work better.
Related: Machine learning, Reinforcement learning from human feedback (RLHF)
- Reinforcement learning from human feedback (RLHF)
A post-training method that uses people’s preferences between possible outputs to help a model produce responses people judge as more useful or appropriate.
Example: Reviewers may compare two chatbot answers, and those comparisons can help train the system to favour answers rated more highly.
Related: Post-training, Reinforcement learning, Model
- Skill
A reusable set of instructions or abilities that helps an AI system carry out a particular kind of task.
Example: A skill might give an assistant a consistent method for summarising a meeting.
Related: Agent, Tool
- Token
A small unit of text a language model processes. A token may be a whole word, part of a word or punctuation.
Example: A short sentence is split into several tokens before a language model processes it.
Related: Large language model (LLM), Prompt
- Tool
A separate service or function an AI system can use, such as searching a website, checking the weather or using a calculator.
Example: A chatbot may use a calculator tool instead of trying to work out a long sum itself.
Related: Agent, Skill
- Training data
The examples used to help a machine-learning system find patterns.
Example: Labelled photographs of plants could be training data for a plant-recognition model.
Related: Machine learning, Model