What AI actually is, underneath the word
It is the most-asked question about the whole subject and almost nobody answers it plainly. Here is what the thing actually does, in one sitting, with no maths.
Today's AI is pattern-matching at enormous scale: it learned from examples instead of being given rules, which is why it is brilliant at things we cannot write rules for and unreliable at things we can.
This is the single most-asked question about AI. It is inside the top five questions asked on Google at all, at roughly 2.2 million searches a month (Exploding Topics). That tells you something worth saying out loud: if you are not sure what it is, you are in an enormous majority.
The old way, and why it stopped working
For most of computing, software did exactly what somebody wrote down. If you wanted a program to spot a cat, you would try to write the rules: fur, four legs, pointy ears.
Try it. Now the cat is behind a chair. Now it is a kitten. Now it is a cat-shaped shadow. The rules multiply forever and the program never gets good, because you cannot write down what makes a cat. You know one when you see one, and you cannot say why.
The new way, in one sentence
So the approach flipped: instead of writing the rules, show the machine millions of examples and let it work out the pattern itself.
Nobody told it what a cat is. It looked at enough cats.
That is machine learning, and it is what nearly everything called "AI" today actually is. The model ends up as a very large pile of numbers describing which patterns tend to go with which.
And then it got weird
The surprising part, and the reason the last few years feel different: do this with language, at enormous scale, and something unexpected happens.
Train a system on a huge amount of text with a boring-sounding job - predict what comes next - and to get good at that job it has to pick up an awful lot on the way. What usually follows a question. How an argument is shaped. What a polite refusal looks like. Which word finishes a joke.
That is a large language model, and it is what sits behind ChatGPT, Gemini, Claude and the rest. Underneath, it is still doing the boring thing. The interesting behaviour is what falling out of doing the boring thing extremely well.
Why it is oddly good and oddly bad
This one idea explains most of what confuses people.
It is good at things nobody could write rules for. Tone. Summarising. Translating. Making something sound less stiff. There was never a rulebook for any of that.
It is unreliable at things with exact answers. It is matching patterns, not looking anything up. A plausible-shaped citation and a real citation look identical from the inside, which is why a model can invent a source with total confidence. It is not lying. There is no moment where it knows the truth and chooses otherwise.
It has no memory between conversations unless a product bolts one on.
It does not know what it does not know. That is the one that catches people. Confidence and correctness are separate things here in a way they are not in a person.
What it is not
It is not a database, a search engine, or a person. It has no goals of its own and nothing that resembles wanting. When output sounds like feeling, that is because it learned from an enormous amount of writing by people who had feelings.
Whether anything more than pattern-matching is going on is genuinely argued over by serious people, and anyone who tells you it is obviously settled - in either direction - is telling you about themselves rather than about the machine.
The one sentence to take away
It learned from examples instead of rules. That is why it can write you a decent birthday message and still get a date wrong, and it is the difference behind almost every other question on this site.
Found something wrong here? That is worth more to me than a compliment. Tell me and it gets corrected on the page, with the date.