You may be using AI every day. Do you know what goes on inside?

Guessing what comes next

How does an AI like ChatGPT figure out what to say?

It guesses the next word. Then the next one. Then the next one.

That’s why, on screen, its answers appear word by word. It doesn’t look up a ready-made answer in a database: it builds one as it goes.

Your turn: pick each word, the way the machine does.

A whole answer, first word to last, is one long chain of bets.

Each word is chosen because it’s likely, not because it’s true.

What are you curious about now?

• Deeper How does it know which word is likely? how it works
• Wider If it’s guessing, can it be wrong? what it changes
• Practical Does the way I ask change the answer? what I do with it

Training

How does it know which word is likely?

It read a vast amount of text and practised guessing the next word. Trillions of times.

Those texts come mostly from the public web, books and articles.

Choose how much it has read, and watch its bets change.

It didn’t learn facts one by one. It learned the habits of language.

Then people fine-tuned it to answer like a polite, helpful assistant.

What are you curious about now?

• Deeper Does it read words the way we do? how it works
• Wider Is it neutral? what it changes
• Practical Where do the words I type go? what I do with it

The pieces

Does it read words the way we do?

No. It cuts text into pieces, and each piece becomes a number.

Specialists call these pieces tokens. It guesses what comes next one token at a time.

Pick a word, or type your own, and watch it get cut up.

The machine never sees letters. Only numbered pieces.

In French, the same meaning often takes more pieces than in English.

What are you curious about now?

• Deeper How can numbers carry meaning? how it works
• Wider Is it neutral? what it changes
• Practical Does the way I ask change the answer? what I do with it

The map of meaning

How can numbers carry meaning?

Each piece becomes a point on a vast map. Words with related meanings end up side by side.

The real map has thousands of directions. We flatten it here so you can see it.

Pick a word to see its neighbours. Then add a new one.

For the machine, meaning is neighbourhood.

Who drew this map? The texts it read.

What are you curious about now?

• Deeper How does it know which “mouse” you mean? how it works
• Wider Is it neutral? what it changes
• Practical Does the way I ask change the answer? what I do with it

Context

How does it know which “mouse” you mean?

Each word looks back at the words before it and uses them to pin down its meaning.

Specialists call this mechanism attention.

Switch sentences and watch “mouse” change meaning.

Context decides meaning. The words you choose matter.

Before each new word, the machine rereads everything before it, including your request.

What are you curious about now?

• Deeper Why does the same question get different answers? how it works
• Wider If it’s guessing, can it be wrong? what it changes
• Practical Does the way I ask change the answer? what I do with it

Measured chance

Why does the same question get different answers?

Because it picks at random from among the likely words. A setting controls how much surprise.

That setting is called temperature.

Set the level of surprise, then roll a few times.

A different answer isn’t truer. It’s just another draw.

If two answers contradict each other, that’s a signal: check.

What are you curious about now?

• Deeper How does it know which word is likely? how it works
• Wider If it’s guessing, can it be wrong? what it changes
• Practical How do I know when I can rely on it? what I do with it

Plausible, not true

If it’s guessing, can it be wrong?

Yes. It aims for what sounds right, not what is true. And it often says true and false things with the same confidence.

These inventions are called hallucinations.

Check the summary: pick every detail that isn’t in the note.

Sounding right isn’t the same as being checked.

The more precise and convenient a detail looks, the more it’s worth checking.

What are you curious about now?

• Deeper How does it know which word is likely? how it works
• Wider Is it neutral? what it changes
• Practical How do I know when I can rely on it? what I do with it

Checking

How do I know when I can rely on it?

Ask yourself three questions: What’s at stake? Where does it come from? How do I check?

Checking doesn’t always take long. It depends on the stakes.

For each use, choose how much checking it needs.

The higher the stakes, the more checking matters. In the end, it’s your name on it.

AI can propose. The decision stays human.

What are you curious about now?

• Deeper Why does the same question get different answers? how it works
• Wider If it’s guessing, can it be wrong? what it changes
• Practical Does the way I ask change the answer? what I do with it

Bias

Is it neutral?

No. It reflects what it read, imbalances included.

A large share of what it read comes from outside Québec, mostly in English.

Change what it read and watch its answer change.

What it’s given to read shapes what it says, and who it speaks for.

The same imbalances show up around gender, culture and jobs. Keep your eyes open.

What are you curious about now?

• Deeper How does it know which word is likely? how it works
• Wider Who builds these machines, and where? what it changes
• Practical How do I know when I can rely on it? what I do with it

Your data

Where do the words I type go?

They leave your device for a company’s servers, often outside Canada. They’re usually kept there, at least for a while.

Depending on the service and your settings, they may also be used to train the next versions of the AI. On several free services, that’s the default.

Before sending, pick what shouldn’t leave.

What you don’t send can’t leak.

In Québec, Law 25 requires organizations to assess the risks before sending personal information outside the province.

What are you curious about now?

• Deeper How does it know which word is likely? how it works
• Wider Who builds these machines, and where? what it changes
• Practical How do I know when I can rely on it? what I do with it

Who builds it

Who builds these machines, and where?

Mostly large companies in the United States and China. But part of the story was written here.

Researchers in Montréal, Toronto and Edmonton laid some of the foundations of modern AI.

And where does your data go? Switch the server’s location, then the company’s home country, and see which rules apply.

Knowing where your data lives, and who controls it, tells you which rules protect it.

In 2017, Canada became the first country with a national AI strategy.

What are you curious about now?

• Deeper How does it know which word is likely? how it works
• Wider Is it neutral? what it changes
• Practical Where do the words I type go? what I do with it

Asking well

Does the way I ask change the answer?

Yes. All it knows about your situation is what you write.

It can’t see your class, your team or your constraints. Unless you describe them.

Add details one by one and watch the answer adjust.

A good request isn’t magic. It’s context.

And useful context almost never needs names.

What are you curious about now?

• Deeper How does it know which “mouse” you mean? how it works
• Wider If it’s guessing, can it be wrong? what it changes
• Practical Where do the words I type go? what I do with it