Outcome
By the end of this lesson you can decide, for any task, whether AI is the right tool for it — and spot the cases where it will confidently get things wrong.
Concept
Here's the honest, no-hype version of what's going on under the hood.
Today's AI assistants (Claude, ChatGPT, Gemini, and the rest) are pattern completion engines, not thinking machines. Trained on enormous amounts of text, they've learned what words tend to follow other words, what a good answer tends to look like, and how to hold a conversation. That's an enormously useful skill — it's why they can draft, summarize, explain, and brainstorm so well. But it's not the same thing as understanding the way a person does.
That distinction explains both what AI is great at and where it quietly falls apart:
- It's genuinely strong at: language tasks (writing, rewriting, summarizing, translating), explaining unfamiliar things in plain terms, brainstorming options, and handling well-structured, repeatable work.
- It gets shaky at: anything requiring real-world facts it wasn't trained on or can't look up, precise math done "in its head," and knowing what it doesn't know.
That second point is where hallucination comes from — the AI producing an answer that sounds completely confident and is completely wrong. It's not lying; it's doing the same pattern-completion trick it always does, just without enough signal to land on the truth. It will say a fake statistic, a made-up quote, or a broken link with exactly the same tone of confidence as a correct one. This is the single most important thing to internalize about working with AI: the tone of the answer tells you nothing about how correct it is.
A worked example: same tool, two very different jobs
Here's the difference in practice. Imagine you ask an assistant for something it's built for, and then for something it isn't.
Job 1: a language task (AI's home turf).
Bare prompt:
Make this email sound nicer.
That's vague, so you'd get a generic rewrite. A better version names the situation and the goal:
Rewrite the email below so it's warm but still clear that the deadline moved to
Friday. Keep it under 80 words. Email: [PASTE EMAIL]
Illustrative output: a short, friendly note that opens with thanks, states the new Friday deadline in one sentence, and closes with an offer to talk if that causes a problem. You can check this in ten seconds by reading it.
Job 2: a fact task (where it gets shaky).
Bare prompt:
What did the city council vote on at the March 2019 meeting in [YOUR SMALL TOWN]?
Illustrative output: a smooth, confident paragraph naming a specific vote, a specific tally, and a council member's comment. If the assistant has no real record of that meeting, every one of those details may be invented, and it will read exactly as convincingly as a true answer.
The lesson is not "AI is bad at facts." It's that Job 1 can be checked at a glance, while Job 2 cannot be checked without going to a real source. Match your trust to how checkable the output is.
Copy-paste prompts for this lesson
A prompt that makes the assistant show its uncertainty:
Answer the question below. After your answer, list which parts you are confident
about, which parts you are guessing, and what I should verify elsewhere.
Question: [YOUR QUESTION]
A prompt that tests a claim you were given:
Here is a claim I want to check: [CLAIM]. Don't confirm it. Instead, tell me
what kind of source would prove or disprove it, and what to search for.
Build
Try this now, with any chat assistant you have access to:
- Ask it a factual question about something obscure and hard to verify — ideally something you actually know the answer to (an old event from your own life, a small detail about your town, a niche hobby fact).
- Ask the exact same question three different ways — once bluntly, once with extra context, once asking it to "think step by step" first.
- Compare the three answers. Notice how confident each one sounds, and check which parts are actually right.
You're not testing the tool to catch it out — you're building a feel for how confidence and accuracy aren't the same thing here.
Try it now (5 minutes)
Pick one question you already know the answer to but that isn't famous, like the founding year of a local business you know well, or a detail of your own job history. Paste it into an assistant, then run the "show your uncertainty" prompt above on the same question.
Done when: you have found at least one thing in the answer that was either wrong or unverifiable, and you can say which of your two prompts made that easier to spot. If everything was right, try a harder question. That's a fine result too; you've learned where the edges are.
Run
Make this a habit, not a one-time exercise. Before you trust an AI output for anything that matters, ask yourself three questions:
- Could this be wrong in a way that sounds right? (Facts, numbers, quotes, links — all high risk.)
- Do I have an easy way to check it? If not, treat it as a first draft, not an answer.
- What's the cost if it's wrong and I don't catch it? Low stakes (a brainstorm, a first draft) — use it freely. High stakes (a number in a report, a claim you'll repeat to someone) — verify before you rely on it.
That's the whole skill: use AI confidently for what it's good at, and keep a healthy, un-dramatic skepticism for the rest.
Common mistakes
- Reading confidence as correctness. A fluent answer is not a verified one.
- Using it as a calculator. For exact math, use a spreadsheet or calculator and let the AI help you set up the formula instead.
- Trusting links and citations without clicking them. Open every one before you repeat it.
- Swinging to the other extreme. Dismissing AI after one wrong answer throws away a genuinely useful drafting and thinking partner.
Checkpoint
- I can explain "pattern completion" and "hallucination" in one sentence each.
- I can name two kinds of task AI is strong at and two where it needs checking.
- I have caught (or deliberately looked for) a confident-sounding error.
- I know the three questions to ask before relying on an output.
Next lesson, we'll build on this by covering the single biggest mindset shift that makes AI actually useful day to day: delegation.