Lesson 004Field guide · AI Basics
Common AI Terms
Prompts, models, tokens, hallucinations — the vocabulary, minus the jargon.
You'll learn
- The seven terms that cover most AI conversations, in plain language
- What a model is and why tools offer you a choice of them
- Tokens and the context window — why long chats lose the thread
- How training data explains both what AI knows and what it invents
- Why you can safely ignore fine-tuning for now
01Section
Seven words, most of the jargon
Articles, settings menus, and product announcements about AI lean on a small set of technical words. Learn seven of them and most of that jargon turns readable: prompt, model, token, context window, training data, hallucination, and fine-tuning.
You already own two. A prompt — the instruction you give an AI tool — you've been writing since Intro to AI. And hallucination you met in What AI Can and Can't Do. So this is five new terms and a deeper look at one you know — each shown doing its job, not listed in a glossary.
What this buys you
Nothing here changes how you type a prompt. It changes what you can read: a settings menu, a headline, a coworker's explanation — without glazing over or nodding along.
Checkpoint
02Section
Model and training data
Plain-language definition
Model — the trained pattern engine behind an AI tool — the thing that actually reads your prompt and generates the response. The chat window is just the counter; the model is the kitchen.
You meet this word every time a tool shows a menu of versions to pick from. That menu is asking which model to use — usually a trade-off between faster-and-lighter and slower-but-more-capable. One model can also power many different products, which is why so many tools feel similar under the hood.
Plain-language definition
Training data — the enormous collection of text a model learned its patterns from before you ever met it. What was in that collection shapes what the model is good at — and what it has simply never seen.
The term in the wild
"The model was trained on data up to early 2025." Translation: its built-in knowledge stops there. Ask about anything newer — prices, versions, events — and it's working from an outdated snapshot or guessing.
Checkpoint
03Section
Tokens and the context window
Plain-language definition
Token — the small chunk of text a model actually reads and writes — roughly three-quarters of a word on average. "Understanding" might be one token; a long sentence, twenty.
You'll meet tokens wherever tools talk about size or price: usage limits, "tokens per month," maximum lengths. It's just the unit AI text is measured in — like kilometers for distance.
Plain-language definition
Context window — the amount of conversation — measured in tokens — the model can see at once. It's the model's working memory: your prompts, its answers, anything you pasted. What fits is visible; what doesn't is gone.
A useful analogy
The context window is a whiteboard. Everything in the conversation gets written on it, and the model can only respond to what's currently on the board. When a long chat fills it up, the oldest notes get wiped to make room — the model isn't ignoring your early instructions; they're no longer on the board.
This explains a frustration you'll hit eventually: deep in a long conversation, the AI seems to forget rules you set at the start. The practical fix is simple — start a fresh chat for each new task, and restate what matters instead of trusting a long thread to remember it.
Checkpoint
04Section
Hallucination, now with the mechanics
What AI Can and Can't Do gave you the definition: hallucination is AI stating something false as if it were fact. With model and training data in your vocabulary, you can now see why it happens.
A model doesn't look answers up in its training data like a database. The data is long gone — what remains are patterns, used to predict likely next words. Usually the likeliest continuation is also true. But when the facts are thin and the pattern is strong, the model produces something that fits the shape of a right answer instead of an actual one. A plausible study name, a realistic-sounding statistic.
Key takeaway
Hallucination isn't the model breaking — it's the model doing exactly what it always does, prediction, on a question that needed lookup. That's why no amount of confidence in the answer counts as evidence.
That's as deep as this path goes. The habit that protects you — verify anything you'll repeat — is next lesson's territory, and the AI Safety & Privacy path covers verification properly.
Checkpoint
05Section
Fine-tuning — and permission to skip it
Plain-language definition
Fine-tuning — extra training that specializes a general model for a narrow job — feeding it many examples from one field until it handles that field's work more reliably.
You'll see this word in marketing: "fine-tuned for legal documents," "fine-tuned on medical research." Now you can translate it — a general model given extra, specialized training.
Here's the part that saves beginners real time: you almost certainly don't need it. Fine-tuning is a project for teams with datasets and budgets. A well-written prompt with good context gets an everyday user most of the benefit at none of the cost. For you, "just use it well" wins.
Permission to skip
If someone insists you need a fine-tuned or custom model to get value from AI, be skeptical. The skills in this path — clear prompts, good context, human review — are the actual entry requirement.
Checkpoint
06Section
A decoder for every term after these
New AI terms will keep arriving. Rather than memorizing ahead, keep a decoder on hand — this prompt turns any jargon-heavy sentence into plain language on demand:
Your jargon decoder — save this
You are my plain-language translator for AI jargon. I'll paste a sentence or short passage from an article, a settings menu, or a product announcement. Rewrite it in everyday language, define each technical term in one line, and end by telling me honestly whether this matters for an everyday AI user or is safe to ignore.
Pause and think: Of the seven terms, which would you have nodded along to last week without really knowing it? That's the one to explain out loud in the exercise.
Checkpoint
Prompt exercise
Teach the terms back
The proof you know a term is being able to explain it simply. Copy this prompt into ChatGPT, Claude, Gemini, Copilot, or whichever AI tool you have access to — the website doesn't run AI itself — and teach the vocabulary back to your tool.
Act as a friendly AI-vocabulary coach. One at a time, in this order, ask me to explain these terms in my own words, as if to a coworker who has never used AI: prompt, model, token, context window, training data, hallucination, fine-tuning. After each answer, tell me what I got right, gently correct anything off, and give one everyday example of the term in use. Keep it encouraging and completely plain-language.
Reflection: Which term took the most effort to explain? Revisit that section once more — if you can explain it simply, you own it.
Quick check
4 quick questions — no pressure
There's no pass or fail here. Answer them all, and we'll show you the answers either way.