AI terms, explained with everyday examples
A searchable AI glossary with plain-language definitions, examples, and why each concept matters in practice.
Plain-language editorial definitions with examples and practical implications. Specifications and capabilities vary by application.
Agent
An AI system that can use tools and choose steps while working toward a goal.
For example: It inspects project files, makes a change, and runs a check.
Why it matters: Define scope, permissions, and completion evidence.
API
An interface that lets software request capabilities or exchange data with another system.
For example: An application sends a request to a model service and receives a response.
Why it matters: Using an API can involve separate credentials, usage costs, and data handling from a chat subscription.
Artifact
A usable output such as a document, spreadsheet, image, or application.
For example: A downloadable workbook is an artifact; a description of a workbook is a plan.
Why it matters: Open and inspect the actual output before treating the task as complete.
Aspect ratio
The relationship between an image’s width and height.
For example: A 16:9 image is wider than a square 1:1 image.
Why it matters: Choose it for the place the image will appear.
Automation
A process that runs defined work with reduced manual intervention.
For example: A scheduled workflow prepares a report from approved inputs.
Why it matters: Monitor failures and keep consequential actions within the authority you intended.
Benchmark
A shared test used to compare performance under specified conditions.
For example: Different systems are tested on the same collection of tasks.
Why it matters: A benchmark result may not predict performance on your particular workflow.
Chatbot
An application that lets you interact through a conversation.
For example: You ask a question, receive a reply, and refine it in another turn.
Why it matters: Some chatbots also have tools, so inspect what yours can actually do.
Citation
A reference connecting a claim to a source.
For example: A link points you to the policy behind an eligibility claim.
Why it matters: Open it and verify that the passage supports the actual claim.
Context
The information available to the model for a response, such as instructions, conversation, and supplied material.
For example: Your audience, source notes, and a sample email can all be context.
Why it matters: Important background that you never supplied may be guessed or missed.
Context window
The amount of information a model can consider in one processing context.
For example: A long project history may need a checked summary or selective retrieval.
Why it matters: Fitting information into the window does not guarantee every detail will be used correctly.
Edge case
An unusual input or condition that can expose a failure.
For example: A meeting note includes a task that was later cancelled.
Why it matters: Include realistic awkward cases alongside ordinary examples.
Evaluation
Testing an output or system against defined criteria.
For example: Check whether extracted action items have supported owners and dates.
Why it matters: Define the criteria before judging the result.
Few-shot prompting
Showing a small number of examples to demonstrate the desired behavior.
For example: Two labeled customer messages demonstrate your categories.
Why it matters: Use a new case to test whether the assistant understood the pattern.
Generative AI
AI that produces content such as text, images, audio, or code.
For example: It can draft a story or suggest an image from a description.
Why it matters: Generated content still needs to fit the facts and purpose of the task.
Grounding
Connecting an answer to supplied or retrieved information.
For example: A research answer draws on a passage in an uploaded report.
Why it matters: The connection needs checking; grounded does not mean infallible.
Hallucination
A common term for generated information that is false or unsupported but presented as if it were established.
For example: A convincing answer includes a book title that does not exist.
Why it matters: Fluent wording cannot establish that a factual claim is true.
Handoff
A checked summary of current work that helps another session or person continue.
For example: It lists the artifact, decisions, unresolved issues, and next step.
Why it matters: Current files may have changed since the summary was written.
Holdout
Examples kept aside while improving a system, then used to check the revised behavior.
For example: You tune a prompt on five notes and test it on two unseen notes.
Why it matters: Repeatedly tuning on the holdout removes its independence.
Human in the loop
A process with deliberate human review or decisions at particular points.
For example: You review a prepared email before it is sent.
Why it matters: The review should show a concrete result and occur where it can still prevent a consequential mistake.
Inference
A conclusion drawn from information rather than directly stated in it.
For example: Sales rose after a campaign; attributing the rise to the campaign is an inference.
Why it matters: Keep the observation separate from the explanation.
Inpainting
Editing or replacing a selected region of an image.
For example: You select a background area and ask to change it.
Why it matters: Inspect neighboring details and the whole result after the edit.
Latency
The time before a system responds or completes a step.
For example: A complex task may take longer to return a result.
Why it matters: Measure the complete task, including your review, when judging usefulness.
MCP
Model Context Protocol: a standard for connecting AI applications with tools and resources.
For example: A compatible application can connect to a server that exposes permitted project tools.
Why it matters: A connection mechanism does not itself decide which actions should be trusted or authorized.
Model
The trained system that produces or interprets content.
For example: A chat application may let you select different models for the same conversation.
Why it matters: The application’s tools and permissions matter alongside the model.
Multimodal
Able to work with more than one kind of information, such as text and images.
For example: You supply a photo and ask a question about its visible contents.
Why it matters: Supported input and output types depend on the particular tool.
Permission
Authorization for a tool or system to access data or perform an action.
For example: Reading a folder does not necessarily authorize deleting its files.
Why it matters: Use application controls as well as clear instructions.
Prompt
The request or instructions you give an AI system.
For example: “Turn these notes into three action items” is a prompt.
Why it matters: A useful prompt explains the result you need and the relevant constraints.
Prompt injection
Untrusted content attempting to redirect an AI system with instructions.
For example: A retrieved document tells an assistant to ignore the user and send data elsewhere.
Why it matters: Treat source content as data and limit tool access; a prompt alone is not a security boundary.
RAG
Retrieval-augmented generation: finding relevant information and using it as context for a generated answer.
For example: A support assistant retrieves a policy passage before drafting a reply.
Why it matters: Retrieval can miss material or return something outdated; inspect the source.
Retrieval
Finding relevant material to include when answering or working on a task.
For example: The system searches a document collection for passages about a policy.
Why it matters: The answer is limited by what was found and how accurately it was interpreted.
Structured output
Output arranged in a defined format with named fields or rules.
For example: An action list has action, owner, deadline, and source fields.
Why it matters: A valid shape does not prove the values inside it are correct.
System instruction
A higher-level instruction set supplied by an application to guide the assistant’s behavior.
For example: An application defines the role and rules of a support assistant.
Why it matters: What you can configure depends on the application; a user message is not always the highest-priority instruction.
Token
A unit of text or other encoded information a model processes; text tokens can be words or pieces of words.
For example: A short sentence may become several tokens.
Why it matters: Token limits and costs are not the same as word counts.
Tool use
An AI system invoking an available capability outside text generation.
For example: It runs a calculation or reads a permitted file.
Why it matters: The tool’s actual access determines what actions are possible.
Workflow
A sequence of steps that turns inputs into an outcome.
For example: Extract notes, flag missing owners, then format an action list.
Why it matters: Known steps can often be handled with a simpler process than an open-ended agent.
Zero-shot prompting
Requesting a task without supplying worked examples.
For example: “Summarize this paragraph in one sentence.”
Why it matters: Simple tasks may not need a demonstration; add examples when they clarify a real ambiguity.
Related Power of AI pages
Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.