Plain-language glossary

Agent: An AI set up to take actions, such as sending an email, booking a meeting or updating a record, not just to write. Give it only the access the job needs.
AI assistant: A chat tool you talk to in plain language, such as ChatGPT, Claude, Gemini or Copilot. Under the hood, each runs one or more models.
Citation: A link or reference the AI gives for where an answer came from. Open it: a citation can point to a page that doesn't say what the AI claims.
Context: Everything the AI can see when it answers: your prompt, files you attached and the earlier messages in the chat.
Context window: The most the AI can take in at once, counted in tokens. In a very long chat or document, the earliest parts can drop out or get less attention.
Deepfake: Audio, video or images made by AI to look or sound like a real person. A voice on the phone asking for a payment can be one, so check through another channel.
Fine-tuning: Training an existing model further on a set of examples, so it gets better at one kind of task or style. Most people never need it: a good prompt and the right documents usually do the job.
Generative AI: AI that makes new text, images, audio or code, as opposed to AI that only sorts or scores things.
Grounding: Tying an answer to a source you trust, such as your documents or a search result, rather than to what the model remembers from training.
Guardrail: A rule enforced by software, such as blocking card numbers before they reach the AI. A policy that relies on people remembering is not a guardrail.
Hallucination: A confident answer that's wrong or made up, such as a fake quote, a wrong date or a source that doesn't exist. It looks exactly like a right answer, which is why you check.
Knowledge cutoff: The date the model's training data ends. Ask about anything after it and the model either searches the web, if it can, or guesses.
Large language model (LLM): The kind of model behind today's AI assistants, trained on huge amounts of text to predict what comes next. That's why it writes fluently, and why fluent doesn't mean true.
Model: The trained system that writes the answer, such as GPT, Claude or Gemini. The app you use is the wrapper around it.
Multimodal: Able to work with more than text: reading images, screenshots, PDFs or audio, and sometimes making them.
Open-weight model: A model you can download and run on your own hardware, so what you type never leaves your machine.
Prompt: What you type to the AI: the question, the instructions and any material you paste in.
Prompt injection: Instructions hidden in a web page, email or file that the AI reads and follows as if you'd written them. It matters most when an agent can act for you.
Reasoning model: A model that works through a problem in steps before it answers. It's slower, and better at maths, planning and tricky logic, but it can still be wrong.
Retrieval (RAG): When the AI looks things up in your documents or on the web before answering, so the answer can quote and cite them. RAG stands for retrieval-augmented generation.
Shadow AI: AI tools people use for work without the organisation knowing or approving them. It usually means the rules haven't caught up, not that anyone is being sneaky.
System prompt: Standing instructions set before your chat begins, by the app or by you, such as custom instructions or a project brief. The model follows them on every message.
Temperature: A setting for how much the answers vary. Low gives steadier, more predictable wording; high gives more variety. Most chat apps set it for you.
Token: The small chunks of text a model reads and writes, often part of a word. Limits and prices are counted in tokens.
Training data: The text, images and code a model learned from. It shapes what the model knows, and the gaps and biases it has.
Training on your data: When a provider uses what you type to improve its models. Many plans let you turn it off, and business plans often do so by default. Check the settings and the terms.

Plain-language glossary

Related