📖 Every AI Word That's Been Confusing You, Explained

Here’s what nobody admits about AI jargon: it’s about three small ideas wearing twenty-nine different hats. There’s a chat window, there’s the brain behind it, and there’s the business of building those brains - and nearly every baffling word you’ve heard belongs to one of the three. Once you see the shelves, the words stop being scary and start being… slightly repetitive, honestly.

This is the fuller dictionary. If you want the starter pack first - the ten words that cover 90% of real life - that’s The Only AI Words You Actually Need To Know, and it’s the better place to begin. This guide is for the day you want the lot: the tool words, the thinking words, and my favourite shelf, the words people mostly use to sound clever in meetings.

No test at the end. Skim it, save it, come back when a word ambushes you.

The Three Shelves

My filing system for every AI term you’ll ever meet. When a new word appears, ask which shelf it lives on, and it’s instantly half as intimidating.

Shelf one: words about the tools. Things you can click. If it describes a button, a setting or a feature, it lives here, and it’s the friendliest shelf - the non-obvious bit is that these words describe things you can just go and try, which beats any definition.

Shelf two: words about how it thinks. What’s happening behind the chat. You never touch these directly, but they explain the tool’s quirks - why it forgets, why it’s sometimes wrong, why long chats go creaky. The non-obvious bit: these are the words that actually make you better at using AI.

Shelf three: words people use to sound clever. Acronyms and grand ideas from the business of building AI. The non-obvious bit, and I say this with love: you will never need to use these words. You only need to recognise them so they can’t be used on you.

Right, let’s stack the shelves.

Shelf one: words about the tools

Chatbot / assistant - the chat window itself, the thing you type into. “Assistant” is the same product after a marketing meeting.

Model - the brain behind the window. Claude, ChatGPT and Gemini are different models with different personalities. You pick one; loyalty is allowed.

Model picker - the little menu in the chat box offering versions of that brain, usually a fast one and a more thorough one. The default is right for almost everyone, which is why it’s the default.

Project - a folder for one area of your life, with its own instructions and files that every chat inside it knows about. The business gets one, the house move gets one.

Artifact - the side panel where Claude builds you an actual thing - a document, a plan, a small working tool - instead of just replying. Grand name, lovely feature.

Voice mode - talking to the AI out loud instead of typing, on the phone apps. Feels odd for about ninety seconds, then feels like the future, especially while driving or cooking.

Connector / integration - a plug between your AI and another tool you use (your calendar, your email), so it can see or do things there with your permission. The word changes by company; the plug idea doesn’t.

Usage limits - the cap on how much you can use a free plan before it asks you to take a breather. On Claude’s free plan the counter resets after a few hours. Nothing’s broken, you’re not banned.

Shelf two: words about how it thinks

Prompt - your message. Any message. The most over-mystified word in the whole building.

Context - everything the AI can currently see: this conversation, anything you’ve pasted. If a detail isn’t in the context, it doesn’t exist as far as the AI’s concerned - which explains most disappointing answers.

Context window - the size limit on that. A very long chat eventually shoves the early stuff off the edge of the table, which is why marathon conversations get forgetful and a fresh chat often perks things up.

Token - the chunks text gets chopped into for reading and writing, roughly syllable-sized. Explains the typewriter effect. Otherwise, safely ignorable.

Memory - notes kept about you between conversations, so tomorrow’s chat knows what Tuesday’s chat learned. Off by default in some tools, free in Claude, and the single biggest “oh, it gets me now” upgrade.

Custom instructions - your standing order: who you are, what you do, how you like answers. Set once in settings, applies forever.

Training / training data - how the model learned in the first place: by reading unthinkable amounts of text before you ever met it. Two everyday consequences: it knows a bit about nearly everything, and it hasn’t necessarily heard this morning’s news (that’s what built-in web search is for).

Hallucination - a confident answer that happens to be false. The model isn’t lying; it genuinely can’t always tell the difference. The fix is boring and works: double-check anything that matters before acting on it.

Reasoning - when a model is set to think for a moment before answering - working through steps rather than blurting. It’s why some answers arrive slower but hold up better, like the difference between a pub answer and a slept-on-it answer.

Shelf three: words people use to sound clever

The fun shelf. Deep breath.

LLM - “large language model”. It means… a model. The thing you’ve been using all along. When someone says “LLMs like Claude”, they’re saying “AI chat tools” with extra letters.

GPT - a technical acronym that escaped containment and became half of a product name. Treat it as a brand, like Hoover. Nobody will ever quiz you on what it stands for, and the people who know don’t bring it up.

Generative AI - AI that makes new stuff (text, images, code) rather than just sorting or recommending existing stuff. Yes, this describes everything in these guides. Mostly a word for conference banners.

Machine learning - the broad family of “computers learning from examples instead of following hand-written rules”. AI’s surname, roughly.

Neural network - the wiring underneath, loosely inspired by how brains connect. You need this word exactly as much as you need “camshaft” to drive to Tesco.

Algorithm - an older word meaning “a recipe a computer follows”. These days deployed to mean “computer thing I distrust”, as in “the algorithm’s hiding my posts”.

AGI - “artificial general intelligence”: a hypothetical future AI as capable as a person at more or less everything. Not a product you can buy; a thing people argue about on podcasts. If someone says it’s definitely five years away, note that it has been definitely five years away for quite a while.

Fine-tuning - taking an existing model and giving it extra training on specific material so it’s sharper at one job. You will hear it misused to mean “I wrote some custom instructions”. Fine-tuning is factory work; instructions are a note to your assistant. The note is what you want anyway.

Open source - models whose inner workings are published for anyone to download, run and modify, versus the big polished commercial ones. A genuinely important idea in the industry; changes nothing about your Tuesday.

Wrapper - slightly sneery word for an app that’s really just one of the big models underneath, in a costume, often with a subscription on top. Worth knowing purely for consumer protection: many “revolutionary AI tools” are a wrapper around the thing you already have free.

Prompt engineering - the craft of asking well. Real, at the professional edges… but for normal humans it means “be specific, give context, say what you want the answer to look like”. You do not need a course in it. You need practice, which is free.

Agentic - the adjective form of “agent”: AI that takes multi-step actions on its own rather than just replying. Currently the most fashionable word in the building, so expect it stapled to products that are, on inspection, a chatbot.

When a word isn’t in here

New jargon breeds faster than any glossary. So here’s the asset that makes this guide permanent - the translate-this-for-me prompt:

Nick this one:

I'm going to paste a sentence or paragraph that's full of AI or tech jargon. Translate it for me:
1. Rewrite it in plain, everyday English - the way a friend would say it.
2. For each piece of jargon, give me a one-line definition and tell me honestly whether it's a word I need to know or just decoration.
3. Tell me the one thing, if anything, in this text that actually matters to a normal person.

Here it is: [PASTE THE JARGON]

Sales page, LinkedIn post, your nephew at Christmas dinner… paste it in, get English back.

Honesty corner

A glossary makes you jargon-proof, not expert. These are working definitions - honest and useful, but a researcher would add footnotes to nearly every one, and the industry will have coined three new words by the time you finish your tea. That’s fine. The Three Shelves absorb new words as fast as anyone invents them, and shelf three absorbs most of them.

If today only has five minutes in it

Save the translate-this prompt. It’s the difference between owning a dictionary and owning a translator, and the translator doesn’t go out of date.

And once you’ve got a taste for turning gobbledegook into English, the same trick works on everything - contracts, bills, error messages, doctor’s letters. That’s the whole of Paste Any Techy Thing In Here And Get English Back. ✨