The words around AI often make simple ideas sound more mysterious than they are. Use this page to translate the terms you will hear in product demos, project meetings, and conversations with technical teams.
An AI system that can work toward a goal by choosing and taking a series of actions. It might search files, use tools, revise its work, and decide what to do next.
In practiceYou ask an agent to compare three attendance reports, find inconsistencies, and draft a summary. It opens the files, checks the figures, and produces the draft.
Keep straightA chatbot usually responds one turn at a time. An agent can take several steps and use tools along the way.
A broad name for computer systems that perform tasks associated with human intelligence, such as recognizing patterns, understanding language, making predictions, or generating content.
In practiceSpam filters, recommendation systems, speech recognition, and ChatGPT all use forms of AI.
A conversational interface that lets you interact with a computer system using ordinary language. The chatbot is the interface, not necessarily the model underneath it.
In practiceChatGPT and Claude are chatbots that provide access to language models and other tools.
The amount of information a model can consider at one time. It includes your instructions, the conversation, attached material, and the model’s response.
In practiceIn a very long conversation, earlier details may fall outside the context window or receive less attention.
Keep straightA large context window does not guarantee that the model will notice or use every detail correctly.
A numerical representation of meaning. Embeddings help a system find items that are conceptually similar even when they do not use the same words.
In practiceA search for “doctor training” might find a document about “physician education” because their meanings are close.
Additional training that adapts a model to examples of a particular task, style, or response format.
In practiceA company might fine-tune a model on approved examples so it follows a specialized classification format.
Keep straightFine-tuning changes the model’s learned behavior. Adding files to a project gives it reference material but does not retrain it.
A large model trained on broad data that can be adapted to many different tasks. Large language models are one kind of foundation model.
In practiceThe same foundation model might help draft an email, analyze a document, or explain a concept.
AI that creates new material, such as text, images, audio, video, or code, based on instructions and examples.
In practiceTurning meeting notes into a first draft of a follow-up email is a generative AI task.
Connecting an AI response to specific source material, such as your documents, a database, or search results, instead of relying only on what the model learned during training.
In practiceAsking the model to summarize an uploaded policy and cite the relevant sections grounds the answer in that policy.
Keep straightGrounding reduces unsupported answers, but you still need to check that the source was read and represented correctly.
A rule, check, permission, or technical limit intended to keep an AI system within acceptable boundaries.
In practiceA guardrail might block personal health information, require approval before sending an email, or flag an answer that lacks a source.
An answer that sounds plausible but contains invented, unsupported, or incorrect information. Some researchers and standards bodies use the word confabulation instead.
In practiceA model may produce a convincing citation to an article that does not exist.
Keep straightA polished tone is not evidence that an answer is true.
The software around a model that gives it instructions, tools, files, memory, permissions, and a way to carry out multi-step work.
In practiceA coding harness might let a model inspect a project, edit files, run tests, and report the result.
Keep straightThe model provides the reasoning and language capability. The harness determines what it can access and do.
A workflow in which a person reviews, approves, corrects, or takes responsibility for important parts of the AI’s work.
In practiceAI drafts learning objectives, but the education team checks their accuracy and approves the final wording.
The process of running a trained model to produce an answer, prediction, or other output. Training builds the model; inference is when you use it.
In practiceEach time a language model responds to your prompt, it is performing inference.
A model trained on very large amounts of text to recognize language patterns and generate useful responses. It predicts likely sequences of tokens rather than looking up a single stored answer.
In practiceAn LLM can draft, summarize, classify, translate, and reason through text-based tasks.
Information an AI product saves and makes available in a later conversation or task. What is remembered, for how long, and who can control it depends on the product.
In practiceA tool might remember your preferred writing style or keep standing instructions for a project.
Keep straightMemory is not the same as the context window. Context is what the model can see now; memory is saved information that may be brought into a future context.
A mathematical system trained to recognize patterns and produce outputs. The model is the underlying engine, while an app such as a chatbot is the product you interact with.
In practiceOne chatbot may let you choose among several models with different speeds and capabilities.
Able to work with more than one kind of information, such as text, images, audio, video, or files.
In practiceA multimodal model might read a chart in a screenshot and discuss it in a voice conversation.
The instruction, question, examples, and background you give an AI system. A prompt can be one sentence or a detailed working brief.
In practice“Turn these notes into a concise email for faculty” is a prompt. Adding the audience, purpose, and source notes makes it more useful.
An instruction hidden in content that tries to make an AI system ignore its intended rules or reveal information it should protect.
In practiceA web page read by an agent might contain hidden text telling it to copy private files somewhere else.
Keep straightPrompt injection is a security problem, not simply a poorly written user prompt.
A method that searches a set of sources for relevant material and supplies the results to a model before it answers.
In practiceAn internal assistant searches the policy library for the most relevant sections, then uses those passages to answer a question.
Keep straightRAG is one common way to ground an answer. It does not guarantee that the retrieved passages are the right ones.
A reusable set of instructions, methods, or resources that teaches an agent how to perform a particular kind of task.
In practiceA presentation skill might tell an agent how to structure slides, apply a house style, and check the finished deck.
Keep straightA skill tells the agent how to work. A tool gives it an action it can take.
A high-priority set of instructions supplied by the product or its builder. It defines the assistant’s role, behavior, limits, and rules.
In practiceA system prompt might require concise answers, prohibit certain actions, and tell the assistant when to ask for approval.
A setting that influences how varied or predictable a model’s output is. Higher values usually allow more variation; lower values usually produce more consistent wording.
In practiceA brainstorming task may benefit from more variation than a fixed-format data classification task.
Keep straightLower temperature does not make an answer more factual.
A small unit of text that a language model processes. A token may be a whole word, part of a word, punctuation, or a short sequence of characters.
In practiceUsage limits and costs are often measured in input and output tokens rather than pages or words.
A capability an AI system can call to take an action or retrieve information outside the model itself.
In practiceWeb search, a calculator, a calendar, file access, and a code runner can all be tools.
Keep straightThe model decides what to say or do. A tool performs a specific action in another system.
The material used to teach a model patterns before it is released for use. It may include text, images, audio, code, or labeled examples, depending on the model.
In practiceA language model learns statistical patterns from training data, but it does not carry a neat, searchable copy of every source.
Try a shorter word or a related idea. For example, try “files” instead of “file retrieval.”