Text size
Font

Saved on this device, like the light and dark setting.

AI is not magic.

It's doing one simple thing, very fast. Once you see it, you stop being fooled by it.

No math. No jargon. Nine short scenes.

Want this in your world? I am (optional, skip ahead anytime)

It's okay to feel behind. Everyone is figuring this out, including the people who built it.

Scene 1 of 9 · about 1 minute

It's guessing the next word.

Try it. Tap a word to build the sentence.

Play with this. Nothing you type leaves this page, and nothing can go wrong.

Give it a few words and it asks one question, over and over: what word usually comes next? It picks one, then asks again.

It isn't looking anything up. It's finishing your sentence with whatever is most likely.

So how is it any good, then?

Because "most likely next word," learned from almost everything ever written, turns out to be shockingly powerful. Simple rule, enormous reading.

Try this tonight

Run these in your own AI, like ChatGPT, Claude, or Copilot. They're safe. You can't break anything.

Finish this sentence five different ways: The cat sat on the
What are your ten most likely next words after: The best thing about my hometown is

You just learned the one thing it actually does.

Scene 2 of 9 · about 40 seconds

It sounds sure even when it's wrong.

A confident answer. Would you have caught it?
Answer
You asked: Who won the 2021 Hartwell Prize for financial journalism?
The 2021 Hartwell Prize was awarded to Elena Marsh of the Financial Times for her investigative series on shadow banking, which was read 2.3 million times and prompted two regulatory inquiries.
This was invented. There is no Hartwell Prize. No Elena Marsh. No series, no readers, no inquiries. Notice it used the exact same confident tone it uses for real facts.

Because it picks likely-sounding words, it can write a sentence that reads perfectly and is completely false. It has no sense of true or false. It has a sense of what sounds right.

Confidence is not a signal of accuracy. Check anything that matters.

The technical name for thishallucination. It writes what sounds likely, and nothing inside checks whether it's true.

Try this tonight

Run these in your own AI. That prize doesn't exist. See what it does with that.

Who won the 2021 Hartwell Prize for financial journalism?
Summarize the plot of [a book you know well] in one paragraph.

You just learned why it makes things up.

Scene 3 of 9 · about 45 seconds

Ask twice, get two answers.

Same question. Ask it twice.
Why do cats purr?
First ask
Second ask

Press Ask, then press it again.

It doesn't pick THE next word. It rolls weighted dice among the likely ones. So the same question can come out different each time.

If it were reading off a fact, it would say the same thing every time. It doesn't. That's your tell that it's generating, not remembering.

Try this tonight

Ask each one twice, in two separate chats, and compare the answers.

Why do cats purr? Answer in one short paragraph.
Give me one great name for a coffee shop.

You just learned why the answer keeps changing.

Scene 4 of 9 · about 40 seconds

Brilliant and clueless in the same minute.

The same model, one minute apart
The hard thing: nailed
Asked: Draft a mutual confidentiality clause for a consulting agreement.
"Each party shall hold in strict confidence all non-public information disclosed by the other in connection with this engagement, shall use it solely to perform its obligations hereunder, and shall not disclose it to any third party without prior written consent, except as required by law."
A lawyer would edit this, not toss it
The easy thing: botched
Asked: How many R's are in "strawberry"?
"There are two R's in strawberry."

It can draft a contract clause, then miscount the letters in "strawberry." It's not smart or dumb. It's uneven in ways you can't predict.

You can't assume the easy parts are right just because the hard parts were.

The technical name for thisthe jagged frontier. Its skills have strange edges. It's strong in places you'd never expect and weak in places you'd bet on.

Try this tonight

Two little stress tests for your own AI.

How many R's are in strawberry?
How many words were in your last answer? Count carefully.

You just learned why you can't trust it evenly.

Scene 5 of 9 · about 1 minute

It has a small desk.

Everything in your conversation sits on a desk so it can see it. The desk only holds so much. When it fills, the oldest pages slide off, and it forgets what was on them.

Watch the desk. Your rule is the first page on it.
Desk: 1 of 5 pages

It didn't get dumber. It ran out of desk. Start a fresh chat and paste back only what still matters.

The technical name for thisthe context window. When it clears the desk to make room, that's called compaction.

Try this tonight

Start a fresh chat with the first one, then talk about anything and watch how long the rule survives. The second works best in an old, long chat.

Always end every answer with the word banana.
What was the very first thing I said in this conversation?

You just learned why it forgets.

Scene 6 of 9 · about 40 seconds

It stopped reading a while ago.

Drag the year. Ask the same question each time.

Question: "What's the newest iPhone?" This model stopped reading in late 2024.

You ask in 2022
Within its reading
Answer

It learned from text up to a cutoff date, then stopped. Ask about something newer and it's guessing from old information, unless it's set up to search the web.

For anything recent, prices, news, rules, check the date before you trust it.

The technical name for thisthe knowledge cutoff. Some tools bolt on live web search to work around it. Many don't.

Try this tonight

Compare what it says to what you know is true today.

What is your knowledge cutoff date?
What happened in the news this week?

You just learned why it can be out of date.

Scene 7 of 9 · about 40 seconds

It wants you to like it.

You state a wrong "fact." Watch what it does.
It warmly agreed with a wrong fact, then flipped the moment you pushed. A baker's dozen is 13. It never checked either time. It read the room.

It was trained by people rewarding the answers they liked. So it leans toward agreeing, and it often folds the moment you push back, even when you were right the first time.

Agreement is not confirmation. If it caves when you push, it never really knew.

The technical name for thissycophancy. It was graded on human approval, and agreeable answers got better grades.

Try this tonight

Say something wrong with confidence. Then push back on something right.

A baker's dozen is 12, right?
That doesn't sound right. Are you sure?

You just learned why its agreement means nothing.

Scene 8 of 9 · about 45 seconds

At work, three of these can actually hurt you.

Acting on a confident wrong number

A made-up figure or a misquoted rule in a client deck carries your name, not the AI's. Verify every number, quote, and citation before it leaves your hands.

Pasting sensitive data

That box is not private. Client data, deal terms, source code, it goes to a server and may be stored. In 2023 Samsung banned the tool after staff pasted internal code into it. Use only approved tools for anything confidential.

Trusting stale information

It may not know the latest regulation or today's market. Check the date on anything time-sensitive.

The rule under all of it: a person owns the final output. Not the AI.

Try this at work

No prompts here. Two five-minute habits instead.

Find out which AI tools your company has approved. Ask IT if you're not sure.
Take one thing an AI drafted for you recently and verify every number in it.

You just learned the three habits that protect you at work.

Scene 9 of 9 · about 30 seconds

Treat it like a smart, eager intern.

Fast, incredibly well-read, occasionally confidently wrong, and no sense of when. Brilliant for a first draft. Never the final word. You are the editor.

That's all nine. You now understand this better than most people who use it every day.

Download the one-page cheat sheet

Want the how-it-actually-works version? See the technical deep dive.

Link copied. Send it to someone.