AI Terms for L&D Teams – A Plain-English Glossary
Someone in a meeting says the tool has "a bigger context window" and "better grounding," and half the room nods while the other half quietly opens a search tab. AI vocabulary moves fast, and much of it was coined by engineers for engineers. You don't need to become one, but you do need enough shared language to ask good questions, set policies and choose tools, which is what this AI glossary is for. Which AI terms do learning and development (L&D) teams actually need to know, and what do they mean in plain English?
L&D teams need a working grasp of about three dozen AI terms, grouped around five ideas: what the tools are, how they produce answers, how to make answers trustworthy, what they can create, and how to use them responsibly. Below, each term gets a short, plain-English definition and, where it helps, a training example. You won't need the mathematics, but you will be able to follow a vendor demo, write a sensible AI policy and explain to a stakeholder why a chatbot gave a wrong answer. Bookmark it and share it with your team so everyone uses the same vocabulary.
The Basics: What the Tools Are
Artificial Intelligence (AI)
A broad term for computer systems that do tasks we usually associate with human thinking, such as understanding language, recognizing images or making predictions. In everyday L&D talk, "AI" now usually means the writing and image tools described below.
Generative AI
AI that creates new content: text, images, audio, video or code. ChatGPT drafting a script and an image tool producing a scene background are both generative AI. The Wikipedia overview of generative AI is a solid primer on the history and common concerns.
Large Language Model (LLM)
The type of model behind text tools like ChatGPT, Claude and Gemini. It learned patterns from huge amounts of text and generates responses by predicting likely next words, which makes it very good with language and less reliable with precise facts it wasn't given.
Model
The trained system that does the work. Products often let you choose between models that trade speed, cost and capability, so think of the product as the car and the model as the engine.
AI Assistant or Chatbot
A conversational interface to a model: you type or speak, and it answers. Microsoft calls its assistants "Copilot," and many people now use "copilot" loosely for any AI helper built into a tool.
Multimodal
A model that can work with more than one kind of input or output, such as text and images together. For example, you might upload a screenshot of a course screen and ask for alt text.
AI Agent
A system that can take a series of actions toward a goal, such as searching, using other tools and updating files, rather than just answering one question. Agents are an emerging area worth watching, and our explainer on AI agents for training shows where they can help L&D teams today.
How the Tools Produce Answers
Prompt
The instruction or question you give the tool. Effective prompts include context, the task, the audience, the format and any limits. "Write quiz questions" is a prompt, while "Write three scenario-based questions for new nurses from the attached policy, one correct answer each" is a better one.
System Prompt or Instructions
Standing instructions that shape every response, usually set by whoever configured the tool. A performance support bot might have instructions like "Answer only from the provided documents and keep answers under 100 words."
Token
The small chunks of text a model reads and writes, often a word or part of a word. Pricing and usage limits are frequently measured in tokens.
Context Window
How much text the model can consider at once, including your prompt, any documents you attach and the conversation so far. A larger context window means you can give it an entire course or a lengthy policy in a single request.
Training Data and Knowledge Cutoff
Training data is the material a model learned from, and the knowledge cutoff is the point after which it has no built-in knowledge. If your policy changed after the cutoff, the model won't know unless you provide the new version.
Temperature
A setting in some tools that controls how varied responses are, where lower values mean more predictable output and higher values mean more creative output. Many chat products don't expose it, but you'll see it in developer tools.
Few-Shot Prompting
Including a few examples of what you want in the prompt, so the model copies the pattern. Paste two finished quiz items in your house style and ask for ten more like them, which is one of the most reliable ways to get consistent output.
Prompt Library
A shared collection of tested prompts for common tasks, which gives a team an easy way to get consistent results.
API
Short for application programming interface, an API is how software communicates with a model directly, without the chat window. Developers use APIs to build AI features into a learning management system (LMS), an authoring workflow or a support tool.
Making Answers Trustworthy
Hallucination
When a model states something false with confidence, like a statistic that doesn't exist or a step that isn't in the procedure. Hallucination is the main reason a human review step belongs in every AI workflow.
Grounding
Giving the model approved source material and instructing it to answer from that material. Grounded answers are more accurate and easier to verify.
Retrieval-Augmented Generation (RAG)
A common method for grounding answers automatically. The system searches your documents for relevant passages, then the model writes its answer from them, which is how most "chat with your policies" tools work.
Embedding and Vector Database
An embedding is a numeric representation of a piece of text's meaning. A vector database stores embeddings so the system can find passages with similar meaning, even when the words differ. This is what lets "angry customer" find a section called "de-escalation."
Fine-Tuning
Training an existing model further on your own examples to change how it behaves. It is useful for some specialized tasks, but for keeping facts current, RAG is usually simpler.
Guardrails
Rules and filters that limit what a tool will do, such as refusing certain topics or always routing safety questions to a person.
Citation
A reference to the source an answer came from, such as a document name and section. In grounded tools, citations let learners and reviewers check the answer in seconds, so treat answers without citations with more caution.
Evaluation
Testing how well an AI tool performs on a set of real questions or tasks with known good answers. Before launching a chatbot, an evaluation with fifty real learner questions tells you far more than a vendor demo.
Human in the Loop
A workflow where a person reviews or approves AI output before it's used. In training, this is simply good professional practice.
What the Tools Can Create
Text-to-Speech (TTS)
Converting written text into spoken audio. Modern TTS from vendors like ElevenLabs sounds natural and makes narration updates as easy as editing the script.
Voice Cloning
Creating a synthetic voice that sounds like a specific person. Always obtain written consent from the person whose voice is used.
AI Avatar Video
Video of a synthetic presenter speaking a script, from tools like Synthesia or HeyGen. It is useful for quick updates and translated versions of talking-head content.
Image Generation
Creating images from a text description. These tools are excellent for scenes and concepts but weaker for exact equipment or diagrams, and you should write alt text based on the image's purpose, not the prompt.
Speech Recognition
Turning spoken audio into text, which powers automatic captions and meeting transcripts. Always review captions for names and technical terminology.
Machine Translation
Automatic translation between languages. Current tools produce a strong first draft, and a fluent reviewer should still check training content before release.
Using AI Responsibly
Bias
Patterns in output that unfairly favor or exclude groups, often inherited from training data. In training content, watch for stereotypes in scenarios and images and for examples that only reflect one kind of learner.
Data Retention and Privacy
What a tool does with what you type or upload: whether it's stored, for how long and whether it's used to improve the model. Enterprise plans often have stricter protections, so check before you upload internal or personal information.
Acceptable Use Policy
Your organization's rules for which AI tools can be used, for what purposes and with which data. If your team doesn't have one, drafting it is an excellent early project.
Disclosure
Telling people when content was created or assisted by AI. A short note in course credits or on a chatbot's welcome screen builds trust, and some internal policies require it.
Which AI Terms Get Mixed Up Most Often?
A few pairs cause most of the confusion in meetings, so here's the quick difference.
| Terms | The difference |
|---|---|
| Grounding vs. fine-tuning | Grounding changes what the model reads; fine-tuning changes how the model behaves |
| Chatbot vs. agent | A chatbot answers; an agent takes a series of actions |
| Model vs. product | The model is the engine; the product is the car built around it |
| Prompt vs. system prompt | Your request vs. the standing instructions behind every answer |
| Hallucination vs. bias | A false statement vs. an unfair pattern across many outputs |
Put This AI Glossary to Work in Your Next Meeting
You don't need every term in this AI glossary memorized. You need your team to share the same meanings, so conversations with vendors, IT, legal and leadership stay clear. A designer who can explain the difference between grounding and fine-tuning, or why a chatbot hallucinated, will make better decisions and earn more trust.
Try this at your next team meeting: pick five terms, ask everyone to explain one in their own words, and give a real example from your work. You'll fill the gaps quickly, and the vocabulary will stick far better than a list ever could.