What is AI, and what it isn’t

“AI” is not one thing. It is a moving target.

Ask five people what artificial intelligence is and you will get five answers: a chatbot, a robot, a recommendation feed, a threat, a spreadsheet with ambitions. None of them are wrong, exactly. They are describing different decades.

So rather than define it in one line, this lesson follows the word — through four attempts at building it, and what survived each time.

Write down enough rules, and thinking will follow.

At a summer workshop in Dartmouth, a small group proposed that every aspect of intelligence could in principle be described so precisely that a machine could simulate it. The plan was elegant: intelligence is logic, logic is rules, and rules can be written down.

For twenty years this was AI. Not statistics — logic.

It worked, too, for a while. Systems that proved theorems, played draughts, and rearranged blocks on a table.

Every rule needed an exception.

Birds fly. Except penguins. Except ostriches. Except a bird with a broken wing, or one in a cage, or a dead one, or a painting of one. Each fix spawned three more cases, and the systems collapsed under their own weight.

Try it yourself — add a few rules and watch what happens to the exceptions.

This is the “knowledge acquisition bottleneck”, and it ended the first AI industry. Twice.

So stop writing rules. Show it examples.

The idea that replaced all of it: do not describe what a cat is. Show the machine ten thousand cats and ten thousand not-cats, and let it work out the rule itself — as a set of numbers nobody has to read.

Click below to place a few examples of your own, then press learn.

Not understanding. Adjusting numbers until the error stops shrinking.

There is no moment of comprehension. There is a measure of how wrong the machine currently is, and a procedure for nudging every internal number a fraction in the direction that makes it slightly less wrong. Repeat a few million times.

That is the whole of it. Everything since 1990 is this idea, applied to bigger models with more data and faster chips.

Stack the idea, and features build themselves.

Put those adjustable numbers in layers, each working on what the last produced, and something unplanned happens: the early layers settle into edge detectors, the middle ones into shapes and textures, the deep ones into wheels, faces, whole objects.

Nobody designed that hierarchy. It is what falls out of the nudging.

This is the “deep” in deep learning — depth of layers, not depth of thought.

Then it turned out bigger just worked.

A new architecture arrived that could be trained on far more data at once. And the field discovered something uncomfortable and enormously consequential: performance kept improving predictably as you added data and compute, long past the point anyone expected it to plateau.

Much of the last decade is that curve, and the money chasing it.

It is still an open question how far the curve goes, and what it does not fix.

From sorting things to producing them.

For most of its history, machine learning answered multiple-choice questions: spam or not, tumour or not, which of these ten thousand words. What changed recently is that the same machinery, at scale, began producing new sequences instead of choosing among old ones.

Same mechanism. Different output. Wildly different public reaction.

The line is blurrier than anyone admits.

Sort these into what you would call AI and what you would not. There are no wrong answers; the disagreement is the point.

Most working definitions today mean something narrow: systems that learn patterns from data and apply them to new inputs. That covers the spam filter and ChatGPT alike, and excludes the thermostat.

It does not want anything.

No goals, no preferences, no inner life, no model of you. When a system produces “I understand how frustrating that must be”, that string was selected because it fits the pattern of the conversation, not because anything was felt.

We are, however, extremely easy to fool. Type something below.

The tendency to read a mind into a responsive surface has a name — the ELIZA effect — and it is far stronger with a modern model than with the toy above.

The moment it works, we stop calling it AI.

Optical character recognition was AI. Chess was the definitive test of machine intelligence — until a machine won, at which point it became “just search”. Spam filtering, route planning, speech recognition, autocomplete: all of it was AI, and all of it is now simply software.

Whatever currently works reliably gets renamed. What remains unsolved keeps the title.

Which means “AI” tends to describe the frontier rather than a technology — and is part of why the term is so slippery in marketing.

A definition you can actually use.

For everything that follows in this course: AI is a system that learns patterns from data, and applies them to inputs it has not seen before.

Not thinking. Not understanding. Not conscious. Pattern-finding, at a scale that makes it feel like more than that — which is both the reason it is genuinely useful and the reason it is so consistently oversold.

Lesson 02 opens up the most famous example and shows you the four steps inside it.

1956 — Logic

Intelligence is rules, written by hand. Elegant, and it broke on the first exception nobody anticipated.

1980s — Expertise

Thousands of rules from human experts. Collapsed under maintenance before it collapsed commercially.

1990s — Learning

Stop writing rules; learn them from examples. Quietly won, and is still the definition that matters.

2012 — Depth

Layers that build their own features. Vision and speech went from unsolved to ordinary within five years.

2017 — Scale

An architecture that rewarded more data and more compute, further than anyone predicted.

Today — Generation

The same machinery producing sequences rather than picking labels. New output, old mechanism.