Course overview
AI Basics for Everyday UsersHow AI Actually Works
A plain-language guide

How AI actually works.

When you chat with AI, it isn't looking up answers in a database. It predicts the next piece of text — one piece at a time — using patterns it learned from a huge amount of writing. Here's exactly what's happening, no maths required.

Tokens Meaning Prediction Temperature Memory
The mental model

It's a prediction engine, not a search engine

Everything else on this page follows from one idea. Get this and the rest clicks into place.

Not a database

It doesn't store pages and look them up. There's no folder it opens to fetch your answer.

Not a person

It doesn't believe, understand or know things the way you do. It has no intentions of its own.

A pattern predictor

It guesses the most likely next piece of text, based on patterns it picked up from huge amounts of writing.

Hold onto this: it predicts what's plausible, not what's verified. That single fact explains why it's fast and fluent — and why it can be confidently wrong.
Where its knowledge comes from

It learned patterns from a lot of text

Before you ever typed a word, the model read an enormous amount of writing and learned which words and ideas tend to go together.

📚

Read a lot of text

Books, articles, code and more

🧩

Learn the patterns

Which words tend to follow which

⚙️

A model that predicts

Keeps the patterns, not the pages

It keeps the patterns, not the original pages. That's why it knows general things well, but doesn't know your private documents, and can't be sure about very recent events.

So when you send a message, four things happen in order. The next sections let you try each one.

Step 1 · How it reads

1Your text becomes tokens

AI doesn't see words the way you do. It breaks text into tokens — small chunks of characters. Type anything below and watch it split.

0
Tokens
0
Words
0
Characters

Notice how common words stay whole, while longer or unusual words split into pieces. Every token takes up room in the model's memory.

Step 2 · How it finds meaning

2Tokens turn into meaning

Each token becomes a set of numbers that capture meaning, so the model can tell that related things are related. Words with similar meanings end up close together. Click any word to light up its nearest neighbours.

Pick a word

Closest in meaning

click a word on the left
Pets Money Time Feelings
This is why a rephrase, a synonym, or even a typo still works — the model lands in the same neighbourhood of meaning.
Step 3 & 4 · The engine

3It predicts the next token — over and over

This is the real engine. Given everything so far, the model ranks which token is most likely to come next, picks one, adds it, and predicts again. Press the button and watch a sentence build itself.

Writing so far
Most likely next tokens

Temperature

Low

The dial decides how it picks from the list above.

PredictableBalancedCreative

Low: it almost always takes the top option. Safer and more consistent — good for policies, factual writing and formal emails.

Watch the bars as you move the dial. Low temperature makes the top option tower over the rest, so output is steady. High temperature flattens them, so it sometimes picks a surprising word — more creative, but riskier.
Its memory

The context window holds only so much

The model can keep just a limited amount of text in mind at once. Keep adding messages and the earliest ones drop out — it can no longer see them.

In the window now0 / 6

Add a few messages and watch the window fill up.

When something matters, start a fresh chat with a clear brief — so nothing you need has quietly fallen out of the window.
Why it makes things up

Confident doesn't mean correct

Because it predicts plausible text rather than looking up verified facts, it can write answers that sound completely sure but aren't true — invented sources, wrong dates, made-up details. The example below looks polished. Reveal what to question.

According to a 2019 Stanford study1, students who receive a follow-up within 48 hours2 are 73%3 more likely to complete their visa paperwork on time. The report, titled “Timely Nudges in Student Administration”4, recommends sending a polite reminder every three days5 until a response is received.
!
A source you can't find — a named study, report or author that may not exist. Always check the source is real.
!
Suspiciously exact numbers — a precise “73%” with no link. Specific-sounding stats are easy to invent.
!
Confident specifics — dates, titles and timeframes stated as fact. Fluent wording is not evidence.
None of this means AI is unreliable junk — it means fluency is not proof. Treat any specific fact, number or source as something to confirm.
Working with it

You're steering the predictions

Since it predicts from whatever you give it, clearer input points it in a better direction. Click a part of the formula to see how each added detail steers the result.

Task + Context + Format + Tone + Constraints

Write an email. Write a polite follow-up email to a student who hasn't shared their visa update. Keep it short, professional, and include one clear action.

This gives only a task. With no context, format, tone or constraints, the model fills the gaps with guesses — so you get something generic.
Click a part of the formula on the left to highlight it in the prompt and see what it steers.
Task: names exactly what to produce, so the goal isn't left to chance.
Context: explains who it's for and why, so the wording fits the situation.
Format: asks for something short, so you don't get a long essay.
Tone: sets a polite, professional voice.
Constraints: requires one clear action, so the email is ready to send.
An honest picture

What it's great at — and what it's not built for

Used for the right jobs it's genuinely powerful. Used as a source of truth, it will let you down.

Great at

Working with language and ideas you give it.

  • Drafting and rewriting emails, notes and documents
  • Summarising long text and explaining ideas simply
  • Brainstorming options and first drafts
  • Changing tone, format or structure of your content
  • Translating and tidying messy text

!Not built for

Anything that needs guaranteed accuracy or live truth.

  • Being the source of truth for facts, law or policy
  • Exact calculations and reliable number-crunching
  • Knowing your private files or very recent events
  • Making the final decision for you
  • Anything where a confident guess could cause harm

The habit that makes AI safe to use: you stay the editor

AI gives you a strong first draft and a thinking partner. The final check is yours.

Check facts and numbersConfirm dates, figures and claims before you rely on them.
Confirm sourcesIf it cites a study, report or link, make sure it actually exists.
Protect confidential dataKeep private and sensitive information in approved tools only.