How AI Works

Introduction

Karthik, a 32-year-old accountant, asked ChatGPT for the capital of Australia. It correctly said Canberra. Then he asked for the population of a tiny village near his hometown — and ChatGPT made up a confident, completely wrong number.

Same tool, very different results. Why? Because of how AI is actually built underneath. Once you understand the four ideas in this lesson, AI's behaviour stops feeling random.

Why this matters

If you know how something works, even at a high level, you can predict when to trust it. Treat AI as a black box and you'll be surprised by every mistake. Understand the four ideas below and you'll start spotting weak answers before they cause problems.

The four ideas behind modern AI

1. Training data

Training data is the giant pile of text, images, or audio that AI learns from. ChatGPT was trained on a huge chunk of the public internet, books, and other text. Image AIs were trained on billions of captioned images. Whatever the AI saw a lot of, it learned. Whatever it didn't see, it doesn't really know.

Analogy: imagine a chef who has read 10 million recipes but never tasted food. They'll write a beautiful recipe — but they're guessing at flavour.

2. Large Language Models (LLMs)

A Large Language Model is the kind of AI behind ChatGPT, Claude, and Gemini. "Large" because it has billions of internal settings. "Language Model" because at its core it's doing one job: predicting the next word, one word at a time, based on everything before it.

That's it. The whole magic is a word-by-word guessing machine that's been trained on so much text that its guesses sound smart.

3. Tokens and context windows

AI doesn't read whole words — it reads chunks called tokens. A token might be a word, part of a word, or even a punctuation mark.

The context window is how much it can "see" at once — your prompt plus its reply. Older models could see roughly a few pages; newer ones can see entire books. When you paste a long document, you're filling its context window.

If you've ever felt the AI "forgot" something you said earlier, it usually means the conversation got long enough to push the earlier part out of the window.

4. Why it sounds so human

It learned from human writing. Books, blogs, forums, articles — all written by humans. So when it predicts word after word, the patterns it picks up are human patterns: tone, humour, structure, politeness, doubt. That's why a good AI reply can feel uncannily like a real person wrote it.

But — and this is important — it has no idea what's true. It has the shape of an answer, not the substance of one.

Step-by-step example: why AI invented a village population

Back to Karthik's village question. Here's what happened, step by step:

1. Karthik's prompt arrived: "What's the population of [tiny village]?"
2. The AI's training data has almost nothing about that village.
3. But the AI is built to always produce a confident-looking answer — it doesn't have a "I don't know" reflex by default.
4. So it predicted a plausible-looking number based on patterns ("villages of this name tend to have a few thousand people").
5. The number was wrong, but the sentence sounded fine.

This is called a hallucination. We'll dig into it in Lesson 1.4.

A prompt that uses this knowledge

Knowing how AI works, Karthik can write a better prompt:

Weak prompt

What is the population of [tiny village name]?

Better prompt

What's the population of [tiny village name]? If you're not sure, say so clearly and suggest where I could check.

Giving the AI permission to say "I don't know" makes it far more honest.

What this knowledge helps you do

• Predict when AI will be strong: common topics, well-known concepts, general writing — lots of training data, great results.
• Predict when AI will be weak: recent events, niche topics, exact numbers, very local information.
• Write better prompts: give the AI context it wouldn't otherwise have (paste the article, the email, the data).
• Know when to use AI search instead of AI chat: if you need facts or current information, use Perplexity or ChatGPT search where it can look things up.

Try it yourself (5 minutes)

Open any chat AI and ask:

In one short paragraph, explain to a beginner what an LLM is and why it sometimes makes things up. Use an analogy.

Compare the AI's analogy to the "chef who's read recipes but never tasted food" one in this lesson. Which felt clearer? Why? Try asking for two more analogies and pick your favourite.

Common beginner mistakes

• Treating AI like a search engine. Search engines look things up. Chat AIs predict from memory. They're different jobs.
• Assuming AI knows current events. Most chat AIs have a "training cutoff" — they don't automatically know what happened last week.
• Pasting huge documents and expecting perfect memory. Even with big context windows, AI handles structure better than tiny details inside long text.
• Not telling AI "say if you don't know." That one small instruction prevents many hallucinations.

Key takeaways

• AI learns from huge amounts of training data and predicts answers based on patterns it saw.
• Large Language Models are word-by-word prediction machines — that's both their magic and their limit.
• Tokens and context windows control how much the AI can "see" at once.
• AI sounds human because it learned from humans — but it doesn't actually know what's true.
• The more you understand the mechanism, the better your prompts and the safer your decisions.