AI glossary

An AI system that works towards a goal over several steps with some independence, deciding what to do next and using tools such as search, files or other apps along the way.

Example: You ask an agent to find train times to another city for Friday and draft a message to your team with two options; it searches, compares and writes the draft for you to review.

Often confused with: A chat assistant answers one message at a time while you steer, whereas an agent carries out a series of actions with less input from you. Because it can act, check what it is allowed to access and review its work before anything is sent, bought or deleted.

The app, website or feature you actually use, such as a chat assistant or a writing helper, which runs one or more AI models behind the scenes.

Example: A note-taking app adds a 'Summarise this page' button: the button and the app are the product, and a model behind them writes the summary.

Often confused with: The product is not the model and not the plan. One product can offer several models or model versions, and what you can do in it depends on the product, your plan and sometimes your region, not only on the model inside.

Learn more: Products, models and plans, Choose a tool and review how it went

A connection that lets software send requests to an AI model and receive answers, so developers can build AI into their own apps and workflows.

Example: An online shop's developer connects to a model through an API so that a first draft of each product description appears when a new item is added, without anyone opening a chat app.

Often confused with: Using a model through an API is different from using the provider's chat app: the features differ, and at some providers API use is billed separately from an app subscription. You do not need an API to use AI tools for everyday tasks.

Learn more: Products, models and plans

Computer systems that do tasks people usually link with human thinking, such as understanding language, recognising images or making predictions.

Example: A photo app that groups pictures by the faces in them uses AI, and so does a chat tool that drafts an email from your notes.

Often confused with: AI is the whole field, not one product. Generative AI is the part that creates new text, images or audio; many other AI systems only sort, rank or predict, such as a spam filter.

Learn more: What AI can and cannot do for you

A fixed set of test tasks used to measure and compare AI models; a score shows how a model did on that specific test under specific conditions, not how useful it will be for your work.

Example: A model can score well on a set of maths competition problems run with particular settings and still write a weak summary of your meeting notes.

Often confused with: A benchmark score is not a general ranking of usefulness. Results depend on the test, the prompts and the settings used, so scores from different sources may not match; trying your own task on two or three tools is a more direct check for your needs.

Learn more: Improve the answer step by step, Choose a tool and review how it went

Source: Artificial Analysis Intelligence Benchmarking Methodology (Artificial Analysis, read 10 Oct 2026)

All the text a model can take into account at one time when writing a reply, including your messages, any material you added and the reply itself.

Example: In a very long chat, the tool may lose track of details you gave at the start if those earlier parts no longer fit in its context window.

Often confused with: The context window is not long-term memory: it covers what is in front of the model right now, while a separate memory feature, where a product has one, stores notes between chats. Its size is measured in tokens, not words or pages.

Learn more: Give the right context

Source: Context windows (Anthropic, read 10 Oct 2026)

A list of numbers that represents a piece of text so that software can measure how related two pieces of text are by how close their numbers are.

Example: A notes app can use embeddings to find a note titled 'car repair costs' when you search for 'how much did fixing the vehicle cost', even though the words are different.

Often confused with: An embedding is not a summary or an answer you read; it is an internal representation used behind the scenes for search, grouping and retrieval-augmented generation.

Source: Vector embeddings (OpenAI, read 10 Oct 2026)

Further training of an existing model on a smaller set of labelled examples so that it behaves in a particular way, such as following a house style.

Example: A support team fine-tunes a model on past customer questions paired with approved answers, so new replies follow the same tone and format.

Often confused with: Fine-tuning changes the model itself and needs prepared data and technical setup. Putting instructions or examples in your prompt, or connecting documents through retrieval-augmented generation, leaves the model unchanged and is usually the simpler option for everyday tasks.

Source: About supervised fine-tuning for Gemini models (Google Cloud, read 10 Oct 2026)

AI that creates new content, such as text, images, audio or code, based on patterns it learned from a very large number of examples.

Example: You ask for five possible titles for a school poster about recycling, and the tool writes five new ones for you.

Often confused with: A search engine points you to pages that already exist, while generative AI writes new text that can sound right and still be wrong. Unless the tool also searches the web, it is not looking anything up.

Learn more: What AI can and cannot do for you

When an AI model produces text that is factually wrong or does not match the material it was given.

Example: You ask for three books on a topic, and one title and author sound real, but the book does not exist when you search a library catalogue.

Often confused with: You cannot spot a hallucination from the tone: a wrong answer can read just as confidently as a correct one. It is also different from out-of-date information caused by the knowledge cutoff, but both need checking against a primary source.

Learn more: How to check what AI tells you, Research with AI and check citations

Source: Reduce hallucinations (Anthropic, read 10 Oct 2026)

The date up to which a model's training data runs, so the model may not know about events or changes after that date.

Example: If you ask about a rule that changed last month, a model without web search may describe the old rule as if it were still current.

Often confused with: The knowledge cutoff is not the date the product was released or last updated; it is usually earlier. When a tool searches the web, its answer depends on the pages it found, so check the dates and sources either way.

Learn more: What AI can and cannot do for you, Research with AI and check citations

Source: How up-to-date is Claude's training data? (Anthropic, read 10 Oct 2026)

A type of AI model trained on a very large amount of text to predict which word or part of a word is likely to come next, which lets it answer questions, write and summarise.

Example: When you type 'Explain photosynthesis to a 10-year-old', the large language model behind the chat app writes the reply a small piece at a time, each piece chosen from what is likely to come next.

Often confused with: A large language model is the engine, not the app. The chat app you open is an AI product that runs one or more models and adds features around them, such as file upload, search or saved chats.

Learn more: What AI can and cannot do for you, Products, models and plans

The specific AI model, identified by a name and usually a version number, that generates the answers inside a product.

Example: A chat app may let you pick from a short list of model names with numbers; asking each one the same question can give answers that differ in length, style and accuracy.

Often confused with: The model version is the engine; the product is the app around it, and the plan decides which versions you can choose. A newer number does not guarantee better results for your task, so try your own task before switching.

Learn more: Products, models and plans

Able to work with more than one kind of input or output, such as text, images, audio or video.

Example: You take a photo of a handwritten shopping list and ask the tool to turn it into a typed list grouped by type of food.

Often confused with: Multimodal does not mean every kind of input works in every product or plan: a tool may read images but not audio, or create images but not video. Check what your tool accepts, and check its description of an image as carefully as any other output.

AI released under terms that let anyone use, study, modify and share it for any purpose, together with the training code, the trained weights and enough information about the training data for a skilled person to build a similar system.

Example: A research group releases a model with its weights, its full training code and a detailed description of where its training data came from, under terms that allow any use, so other teams can inspect it and retrain their own version.

Often confused with: Open source AI and open weights are different labels. A model whose weights you can download, but whose training code or data information is not shared, or whose licence restricts how you may use it, is open weights but not open source AI under the Open Source Initiative definition. Read the licence rather than trusting the label.

Source: The Open Source AI Definition – 1.0 (Open Source Initiative, read 10 Oct 2026)

A model whose trained weights, the numbers it learned during training, are published so people can download, run and adapt it themselves under the terms of its licence.

Example: A developer downloads an open-weights model and runs it on their own computer, so documents stay on that machine, but they still have to follow the licence that comes with the model.

Often confused with: Open weights is not the same as open source AI. The licence may limit commercial use or certain purposes, and the training data and training code may not be released at all; open source AI, as the Open Source Initiative defines it, requires the freedoms to use, study, modify and share plus data information, code and weights.

Source: Gemma 4 model overview (Google, read 10 Oct 2026)

The level of access you have to an AI product, such as a free tier or a paid subscription, which affects which features you get.

Example: Two classmates use the same chat app, but one is on the free tier and one pays monthly, so they may see different features, usage limits or model choices.

Often confused with: A plan is not a model or a product; it is the account level you sign up for. At some providers, the app subscription and developer API use are billed separately.

Learn more: Products, models and plans, Choose a tool and review how it went

Source: I have a paid Claude subscription (Pro, Max, Team, or Enterprise plans). Why is Claude API usage billed separately from my paid Claude plan? (Anthropic, read 10 Oct 2026)

The options in an AI product that control what happens to your data, such as whether your chats may be used to improve the provider's models and whether your chat history is kept.

Example: Before using a chat tool for work, you open its settings, find the option about using your chats to improve models, and choose what fits your situation and your organisation's rules.

Often confused with: Privacy settings differ by product, plan and account type, and they can change, so check the current settings yourself instead of assuming a default. Turning off model improvement does not necessarily mean nothing is stored, so still avoid pasting sensitive personal, company or other people's data.

Learn more: Choose a tool and review how it went, Break down an office task

Source: Data controls in ChatGPT (OpenAI, read 10 Oct 2026)

The instruction, question or material you give an AI tool to tell it what you want.

Example: 'Rewrite this paragraph for a 12-year-old reader in under 80 words' is a prompt, and so is a pasted email followed by 'List the three things I need to do'.

Often confused with: A prompt does not need special keywords or tricks. Plain language is enough: say the task, give some context and say what form you want the answer in.

Learn more: Asking your first question, Start with a clear goal

A way of working in which a model takes extra steps to break a problem down before giving its final answer, often used for maths, logic, planning or code.

Example: Asked to plan a weekly study timetable around two exams and a part-time job, a model in reasoning mode may take longer and work through the constraints before showing the final plan.

Often confused with: Reasoning is not the same as being right: the steps can look sensible and still contain a mistake, so check the result. It is usually slower and may use more of your usage allowance, so a quick rewrite or summary may not need it.

A method in which a system first finds relevant passages in a chosen set of documents and passes them to the model, so that its answer is grounded in that content.

Example: A company help bot searches the staff handbook for the section on holiday leave, then writes its answer from that section and can show which section it used.

Often confused with: Retrieval-augmented generation does not retrain the model; it adds looked-up text to the request each time, which is how it differs from fine-tuning. The answer can still be wrong if the wrong passage is found or misread, so open the passage it relied on.

Source: Retrieval-augmented generation (RAG) in Azure AI Search (Microsoft, read 10 Oct 2026)

A small unit of text that a model reads and writes, which can be a single character, part of a word, a whole word or a punctuation mark.

Example: A short, common word is often a single token, while a long or unusual word may be split into several tokens.

Often confused with: A token is not the same as a word, so limits and prices stated in tokens do not convert exactly into words or pages. Different models split text in different ways, so the same text can count as a different number of tokens.

Learn more: Give the right context

Source: Understanding and counting tokens (OpenAI, read 10 Oct 2026)

The ability of an AI model to call outside functions or services, such as a web search, a calculator or a calendar, and use the results in its answer.

Example: When you ask for the opening hours of a local museum, a tool with web search can look up a current page, while a model without tools can only answer from what it learned in training.

Often confused with: Tool use is a single ability; an agent is a system that may use many tools over many steps. A model without tools cannot check the web or your calendar, even if its answer sounds as if it did.

Learn more: What AI can and cannot do for you

Source: Tool use with Claude (Anthropic, read 10 Oct 2026)

The process in which an AI model learns patterns from a very large amount of example data, which shapes how it responds later.

Example: A language model is trained on huge collections of text, so it picks up how sentences are usually built and which words and facts tend to appear together.

Often confused with: Training happens before you use the model; when you chat, it works from what it already learned plus what is in the conversation. Your chats only become training data if the provider uses them that way, which privacy settings may let you control.

Learn more: What AI can and cannot do for you