AI vocabulary: the 20 terms you need to sound fluent

Most AI terms are simple ideas wearing a lab coat. A model is a program that learned patterns from examples. A prompt is what you type. A token is a chunk of a word. A hallucination is a confident wrong answer. Learn about 20 of these words and the entire conversation, the vendor pitches, the podcasts, the LinkedIn posts, stops sounding like a foreign language. That is the whole job of this post.

This is part 2 of a six-part series for total beginners. Part 1 gave you the plain-English mental model of what AI is. This one hands you the vocabulary so you can follow any AI conversation without nodding along blindly.

What are the AI terms I really need to know?

You do not need 200 terms. You need about 20, and they cluster into three groups: what the thing is made of, what happens when you talk to it, and how it gets better. I call it the 3-bucket map, because once a new word lands in a bucket, you already half-understand it.

Keep those three buckets in your head. Every buzzword you hear this year will drop into one of them.

Bucket 1: the ingredients

These five words describe what an AI tool is built from.

A model is the trained program itself, the thing that does the predicting. When someone says "which model are you using," they mean which specific brain, like GPT-5 or Claude or Gemini.

An LLM, or large language model, is a model trained specifically on text. The chatbots you know are LLMs. You will hear it like this: "It is just an LLM, so it is great at writing and bad at math."

Training data is the pile of examples the model learned from, usually a huge chunk of the public internet plus books and articles. You will hear it like this: "The model has a cutoff date because its training data stops there."

Parameters are the internal dials the model tunes while learning, and there are billions of them. You do not touch these. You will hear it like this: "It is a 70-billion-parameter model," which is a rough proxy for how big and capable it is.

A token is a chunk of text the model reads and writes in, usually about three-quarters of a word. Tools count usage and set limits in tokens. If you want to keep costs and length in check later, understanding tokens saves you real money.

Bucket 2: the conversation

These are the words for what happens when you use the tool.

A prompt is the instruction you type. "Write me a follow-up email" is a prompt. The better your prompt, the better the output, which is the entire reason learning to write prompts is worth an afternoon.

The context window is how much text the model can hold in mind at once, your prompt plus its reply plus anything you pasted. You will hear it like this: "I hit the context window, so it forgot the top of the document."

Inference is the fancy word for the model running and generating an answer. You will hear it like this: "Inference costs go up when a lot of people use the tool at once."

The output (sometimes called the completion) is what the model gives back. Simple.

A system prompt is a hidden instruction that sits above your chat and shapes the tool's behavior and personality. You will hear it like this: "Their custom bot has a system prompt telling it to always answer as a friendly accountant."

Temperature is a setting for how predictable versus creative the answers are. Low temperature gives safe, repetitive answers. High temperature gives wild, varied ones. Most chat apps hide this, but you will hear developers mention it.

A hallucination is when the model states something false with total confidence, like inventing a citation or a statistic. This is the single most important term in the bucket. The model predicts what a right answer looks like, not what is true, so it sometimes produces polished nonsense. Part 3 explains exactly why this happens.

Bucket 3: the upgrades

These words describe ways people make a base model more useful.

You do not need to build any of these. You need to recognize them when a tool, a vendor, or a course puts them on a slide. An agent is worth flagging: it is the word behind most of 2026's hype, and it just means an AI that can act, not only answer. The rest of this series builds up to that.

What does knowing these words change in practice?

Take a composite. A shop owner I'll call Priya sat through a demo where the salesperson said their tool "used RAG on your data with guardrails and a fine-tuned model." Six months earlier that sentence would have made her feel behind, so she would have nodded and signed. This time she heard three plain ideas: it can read her documents, it has safety rules, and it was trained on examples like hers. She asked three sharp follow-up questions instead of one vague one, and she negotiated a better trial.

That is the real return on vocabulary. It moves you from nodding to negotiating.

Jargon is a tax on people who never learned the words. This post is your refund.

Do this next

Open any chatbot and paste this: "Explain the difference between the context window and training data using my coffee shop as the example." Watching the tool use these words on your own business locks them in faster than any glossary. Inside the WorkSmart OS, the monthly AI trainings keep this vocabulary current as new terms show up, so you are never the person nodding along in the demo.

Next up, part 3 shows you how these chatbots really work, with no math and no mysticism.