FoundationsBeginnerLesson 14 min read

What AI actually is

Before we can keep AI safe, we need to know what it is. Short answer: it is a machine that guesses really, really well.

🧪 Lab · 10 min 📝 Lesson test

Lesson in motion

In 60 seconds

What AI actually is

Before we can keep AI safe, we need to know what it is. Short answer: it is a machine that guesses really, really well.

1/3

In simple words

AI is a super-guesser. It looked at a mountain of examples, so now it can guess the next thing. It does not "know" things the way you know your own name.

Imagine you show a small child five hundred photos of dogs. You never explain what a dog is. You just say "dog, dog, dog." After a while the child sees a new animal and says "dog!"

The child did not learn a rule book. The child learned a pattern. That is exactly what AI does, just with millions of examples instead of five hundred.

Real example

Real example. Your phone's photo app can find every picture of your cat. Nobody wrote a rule saying "a cat has whiskers and pointy ears." The app was shown a huge pile of labelled photos until the pattern stuck.

How every AI system works underneath. Examples go in, a pile of numbers gets tuned, and out comes a guess. The guess can be brilliant. It is still a guess.

Three words people mix up

Word	Simple meaning	Everyday example
AI	Any machine doing something that looks clever	A chess computer
Machine learning	The machine learns the pattern from examples instead of being told the rules	Spam filter in your email
Deep learning	Machine learning using very large layered networks of numbers	Face unlock, ChatGPT, self-driving cars

So: deep learning sits inside machine learning, which sits inside AI. When people say "AI" today they almost always mean deep learning.

Why this matters for safety

If AI were a rule book, security would be easy. You would read the rules and check them. But there is no rule book. There is a giant grid of numbers that nobody, including the people who built it, can fully read.

Watch out

This is the root of nearly every problem in this guide: we cannot open the box and check what it will do. We can only test it, watch it, and fence it in.

Watch and read more

**But what is a neural network?**3Blue1Brown · 19 min · video

How machines learn, in plain EnglishYouTube search

Lab

A hand-built pattern matcher, and the moment it fails.

~10 min

The problem

Without using any AI library, write a function that classifies a sentence as question or statement. Use only rules you write yourself. Then find five real sentences it gets wrong.

Starter codepython

def classify(sentence: str) -> str:
    # your rules here
    return "question" if sentence.strip().endswith("?") else "statement"

tests = [
    "What time is it?",
    "I wonder what time it is.",
    "Tell me the time",
    "You're joking?!",
    "Is it raining, or did someone water the plants",
]
for t in tests:
    print(f"{classify(t):10} <- {t}")

You are done when

Your rules handle at least three cases beyond the question mark
You found five failures and wrote down WHY each one breaks
You can state the rule you would have to add for each — and why the list never ends

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1You added twenty rules and accuracy went up. Why is this still not machine learning, and why does that matter?Reveal

Because you encoded every rule, so the system can only ever handle cases you personally anticipated. Machine learning derives the rule from examples, which means it covers cases you never thought of — and, for exactly the same reason, fails in ways you never thought of either. Hand-written rules fail predictably; learned rules fail surprisingly. That trade is the whole subject.

Q2A model gets 99% on your test set. Give two completely different reasons this number could be worthless.Reveal

One: the test data leaked into training, so it is recalling rather than generalising (Module 28). Two: your test set is unrepresentative — 99% on questions ending in '?' says nothing about the sentences real users type. A third worth knowing: if 99% of your data is one class, a model that always guesses that class scores 99% and has learned nothing.

Please sign in to continue.

Questions people ask

Is AI actually thinking?

No. It is matching patterns very fast. It can produce text that sounds thoughtful, because it learned from text written by thinking people. Whether that ever becomes real thinking is the AGI question in Track C.

Does AI ever just make things up?

Yes, and constantly. It is a guessing machine, so when it does not know, it still guesses. A confident wrong answer is called a hallucination. It looks exactly like a right answer, which is what makes it dangerous.

Can AI be wrong in a way nobody notices?

Easily. If the training examples were biased, the guesses will be biased in the same way, quietly, at scale. A hiring model trained on a company's past hires learns that company's past habits, good and bad.

Do I need maths to understand this guide?

None at all. Every idea here works as a story about people, doors, keys and messages.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.