Lesson 1 · 6 min
What AI Actually Is
A plain-language picture of the tool. No math required.
When you talk to a modern AI assistant — ChatGPT, Claude, Gemini, or another — you are talking to a kind of program that has read an enormous amount of text and learned, deeply, what words tend to come next in any given context. When you write to it, it predicts what an excellent reply would look like, one piece at a time, until the reply is finished.
That sounds underwhelming. But prediction at this scale turns out to do something close to thinking. To predict what should come next in a thoughtful response, you have to model what a thoughtful response sounds like.
Ask it to describe itself
The prompt
In two short paragraphs and plain English, explain what you are and one thing you're genuinely bad at.
What you might get back
I'm a program trained on an enormous amount of text. I don't 'know' facts the way a database does — instead, I predict the most likely next words to build a helpful response, one piece at a time. That's why I can sound fluent and thoughtful across almost any topic. One thing I'm genuinely bad at: knowing what I don't know. I can state something false with total confidence, because to me a wrong answer and a right one can look equally plausible. So I'm a great first draft and a poor final authority — double-check anything that matters.
This is exactly your task below. Notice how asking for 'two short paragraphs' and 'plain English' produced a clear, honest answer instead of a wall of text.
Why this works — fluent is not the same as knowledgeable
Hold this one fact and most of AI's behavior stops being surprising: the model isn't looking anything up. It is predicting the next stretch of text that would plausibly follow yours, word by word — that's not a feature bolted on, it's the whole of what it does. Fluency is what it was built for; being right is only a side effect of the true text it happened to learn from.
That gap — fluent, but not necessarily knowledgeable — is why it can sound most certain exactly when it's inventing. The same instinct transfers everywhere you'll ever use it: a smooth, confident answer is evidence of good prediction, never proof of a checked fact.
Why it sometimes makes things up
Because the model is predicting what good text looks like — not retrieving facts from a database — it can sometimes invent things that sound completely plausible but are wrong. This is called a hallucination. It does not know it is doing this. Knowing this failure mode is what lets you use the tool wisely (we'll spend a whole track on it later).
A confident hallucination
The prompt
What year did the novelist Marcus Vale win the Pulitzer for 'The Salt Road'?
What you might get back
Marcus Vale won the 1998 Pulitzer Prize for Fiction for 'The Salt Road,' a sweeping novel about a family's journey across the American Southwest.
Marcus Vale and 'The Salt Road' don't exist — but the question assumed they did, so the AI confidently invented a detailed answer. That's a hallucination. The habit that protects you: ask it to flag what it's unsure about, and verify anything that matters.
Why the same question gives different answers
If you ask the same question twice, you may get slightly different replies. This is by design — a small dose of randomness is why responses feel alive instead of robotic. It also means that if a reply isn't great, asking again can help.
How to work with a predictor
- Give it context — it walks into every chat knowing nothing about you.
- Treat it as a brilliant first draft, not a final authority. Verify anything that matters.
- If an answer is off, just ask again or push back — variation is normal, and a second try often helps.
- Ask it to flag when it's unsure, so confident-sounding guesses are easier to catch.
Recap
- It predicts likely words — it doesn't look facts up in a database.
- By default it has no memory across chats, no files, no live internet.
- It can be confidently wrong — that's a hallucination.
- You steer the quality with context, and protect yourself by verifying what matters.
Open any chatbot. Ask it to explain, in plain language, what it actually is — and one thing it is genuinely bad at. Ask for two short paragraphs.
Then bring back your prompt and what it said. (This is just to break the seal — there's no wrong answer here.)
Example: "In two short paragraphs and plain English, explain what you are and one thing you're genuinely bad at."
Tool: any chatbot (ChatGPT, Claude, Gemini, or Copilot)
Your work saves automatically when you leave — come back any time to pick up where you left off.