AI terms explained: a plain-English glossary for beginners

The five words that matter in your first hour

Most writing about AI assumes a vocabulary nobody ever taught you. The useful part is small. Five words carry nearly all the practical weight, and the rest can wait months.

A model is the AI itself: the trained system that reads your words and produces an answer. The app is not the model. ChatGPT is an app running OpenAI's models, Gemini is an app running Google's, and Whizi is an app that holds several companies' models at once. This distinction earns its keep the first time an answer disappoints you, because switching the model is usually a bigger lever than rewording the question.

A prompt is just what you type. There is no secret syntax, no command list, nothing to memorize. A plain request that names who the output is for, what job it has to do, and what shape you want back will beat a clever one-liner nearly every time.

A chatbot is the conversation window you type into. Everything you have said inside one conversation stays visible to the model while it answers, which is why follow-ups like "make it shorter" or "more formal" work without you repeating the original request. Open a new chat and that shared memory is gone, which is what you want when the topic changes: leftover context from an unrelated question is a common source of answers that feel slightly off.

A token is a chunk of text, roughly three quarters of an average English word. Models read, count, and bill in tokens rather than words, so a 1,000-word document lands somewhere near 1,300 tokens. You will mostly meet the word on pricing pages and usage limits.

A context window is how much text a model can hold in view at once, measured in tokens. It has to cover the document you pasted, the whole conversation so far, and the answer being written. When you drop a fifty-page contract into a chat and the replies start quietly contradicting page three, you have found the edge of it. What a context window is has the real numbers; the working rule is short. One long conversation about one thing, then start fresh.

Something to do with that right now: open any AI app and ask it, in your own words, to explain a term you saw this week and did not recognize. Then ask a second question, "What did you assume I already know?" That follow-up turns a definition into an explanation, and it works on any topic.

Check the answer against something else, though. These systems state wrong things in exactly the same calm voice they use for right ones, and definitions are a soft target: a plausible invented detail slides past you because you had no way to catch it. That failure has a name, and it is in the table below.

The rest of the glossary, and whether you need it yet

These are the words waiting for you in headlines, pricing pages, and comment sections. The third column is the honest part. Most of this is not day-one knowledge, and pretending otherwise is how beginners end up feeling stupid about something that does not affect them.

Term	Plain meaning	Do you need this?
LLM	Large language model, the kind of model behind every chat app you have heard of. When an article says LLM, read it as the AI	Yes, so headlines stop reading as noise
Generative AI	Any AI that produces new text, images, audio, or video instead of sorting things that already exist	Yes, one sentence is enough
Training	The long, expensive process of feeding a model enormous amounts of text so it learns patterns. It finished before you opened the app	Yes, briefly
Training cutoff	The date the training data stops. Ask about anything later and the model is guessing, unless the app searches the web	Yes, this one catches people out
Hallucination	A confident, fluent, completely invented answer. The flaw that matters most	Yes, on day one
Multimodal	The model handles more than text: photos, PDFs, audio, sometimes video	Yes, the moment you want to upload a picture
Fine-tuning	Retraining an existing model on your own data so it specializes	Not yet, this is a company project
Temperature	A dial for how random the wording is. Low is predictable, high is loose	Not yet, sane defaults are chosen for you
Agent	An AI that takes several steps on its own instead of replying once	Not yet, and see the next section
API	The plumbing developers use to wire a model into their own software	Not yet, unless you write code
RAG	Retrieval augmented generation: the model looks things up in a document store before answering	Not yet, though any app that answers from your uploaded files runs on it
Open source model	A model whose weights anyone can download and run, such as Llama or DeepSeek	Not yet, it changes little for chat users
Reasoning model	A model that works through a problem in visible steps before answering. Slower, better at math and logic	Soon, the first time a hard question fails
System prompt	Hidden instructions the app gives the model before you type anything	Not yet, useful when you build your own assistant

Two rows deserve a footnote. Hallucination is not an occasional glitch to be patched away; it falls out of how these systems generate text, which is why why AI gets things wrong is a guide of its own rather than a bullet point. And multimodal is worth ten minutes as soon as you have a photo of a receipt, a screenshot of an error message, or a scanned form to deal with. Working with images and documents shows what that looks like on a normal Tuesday.

Words marketing uses that mean less than they sound like

Vendors are not lying, exactly. They are using words that once meant something specific and now mostly mean "we shipped something". Knowing the gap is what stops you paying an upgrade fee for a dropdown of canned replies.

• Powered by AI. Usually means one feature calls a model built by somebody else. Sometimes it means a suggestion box you can switch off. Worth asking: which model, and is it on by default?
• Smart. Predates this whole wave and now means very little. A smart folder is a saved search. A smart reply is three canned phrases. Look past the adjective for a verb: what does it do to your work?
• Next-generation. A claim in the future tense. It is perfectly compatible with a product being worse than the version it replaced, and it tends to appear when there is no measurable improvement anyone can name.
• AI agent. The hardest-working phrase of 2026. Used strictly, it means software that plans and carries out several steps by itself, like booking a trip end to end. Used in marketing, it usually means a chatbot with a friendlier avatar. Ask what it does when it gets stuck, and whether it can act without checking with you first.
• AI-first. A statement about internal strategy, not about your experience of the product. It tells you where a company points its engineering budget, which may not yet have reached the feature you were about to pay for.

One question cuts through all of it: ask what the feature does when it is wrong. Products with real AI inside have an answer, and it involves review steps, an undo, a confidence signal, or a visible source. Products carrying a sticker go quiet.

Why model names look like license plates

A model name has two parts doing two different jobs. The word is the family, the product line a company keeps developing: GPT from OpenAI, Claude from Anthropic, Gemini from Google, Llama from Meta. The numbers and the suffixes are the version, and often the size, so one family will ship a small quick member and a large careful one on the same morning. Reading a name is mostly a matter of separating those halves and ignoring the marketing in between.

You may have noticed this article does not name a single version number. That is deliberate. Versions change every few weeks, and any glossary that pins them down is wrong by the next season. What does not change is the shape of the choice. Small models are fast and cheap and completely fine for tidying an email. Large ones are slower and better with long documents and hard reasoning. Each family also has a personality that shows up in tone and in how strictly it sticks to a source you gave it.

The newest release is not automatically right for your task either. Newer usually means better benchmark scores, and benchmarks measure exam-style problems rather than whether the writing sounds like you. Plenty of people find that an older, cheaper model writes their emails better, and they save the expensive one for the two hard jobs a week. You only learn this by running the same real task through two or three models and reading the outputs next to each other, which is what choosing an AI model walks through step by step. Doing that without paying three providers is the whole reason Whizi puts GPT, Claude, Gemini, and image models inside one subscription.

Five terms worth learning at the three-month mark

By month three you will have walked into a few walls. These five words are the names of those walls, which is why they are worth learning then and not now.

1. Reasoning model. Starts mattering the first time a model confidently botches a multi-step calculation or a logic puzzle. Switching to a reasoning model often fixes it outright, at the cost of waiting longer for the answer.
2. System prompt. Matters when you notice you are pasting the same background paragraph into every new chat. Saving it once as a standing instruction is the fix.
3. RAG. Matters when you want answers grounded in your own files rather than the open internet, for example a folder of policies nobody has read since 2019.
4. Temperature. Matters when output feels either too samey across attempts or too erratic, and you would rather turn a dial than keep rerolling.
5. Fine-tuning. Matters almost never for an individual, and it is worth knowing why: it costs real money, needs hundreds of clean examples, and a good prompt on a good model beats it for nearly every use people imagine it for.

If you want the layer underneath all of this, what AI actually is covers the mechanism without the math, and how to talk to AI turns that prompt paragraph from the first section into a habit. Leave this page open in a tab for a week. The vocabulary stops being intimidating around the fourth time you meet a word you already know.