What people mean when they say AI

There are two ways to make a piece of software do something. A person can write out the rules for it to follow, or a person can collect examples of the job done right and let the software work out its own rules. Nearly everything that has ever run on a computer is the first kind. The word AI, wherever you met it this week, is pointing at the second.

Written by hand, learned from examples

Tax software applies the tax code exactly as programmers typed it in. Your alarm goes off because a line of code says: at seven, ring. Every behaviour of software like this was decided by a person, in advance, on purpose. When it misbehaves, there is a wrong rule somewhere, and a person can find the rule and fix it.

The spam filter on your email is the other kind. Nobody wrote down rules for what junk mail looks like. Nobody could say what those rules are, exactly, and the junk keeps changing anyway. Instead the filter was shown millions of messages that people had marked as junk, and millions that people had kept, and it settled into its own way of telling one from the other. Showing software examples until the behaviour appears is called training, and it is what the word AI now reliably signals: behaviour learned from examples rather than written by hand.

The word itself is older and baggier than that. Across seventy years it has been pinned on plenty of software whose rules were entirely hand-written, and how it wandered is a story of its own. But on a product or in a headline today, AI means trained. That is the live distinction under the label, and it is worth more than the label is.

You were using it before it could talk

List the trained software you touched today and the chatbot arrives late.

A trained system intercepted your junk mail before breakfast. Every feed you scrolled was arranged by one: the posts appear in the order software predicts will keep you scrolling, an order learned from what millions of people lingered over before you. The voice note your phone turned into text, and the automatic captions under the video, came from software trained on oceans of recorded speech. Type "beach" into your photo app and it finds the beach, though you never labelled a photo in your life. Even the keyboard guessing your next word is a trained system, so small and so domesticated that nobody has called it AI in years. Different jobs, one word over all of them. AI is an umbrella, not the name of any single machine.

Then there is the pair everyone means now: the chatbot and the image generator. They belong on the same map as the rest, and what sets them apart is the output. The older systems judge and sort, and what they produce is small — a ranking, a label, a transcript — and easy to mistake for the ordinary working of the machine. The new pair make things: paragraphs, pictures. AI did not arrive in your life on the day it started talking to you. That is just the day it stopped being easy to miss.

The label does not stay put

In 1997 a chess machine beat the reigning world champion, and for a while Deep Blue was the most famous artificial intelligence on earth. Chess programs far stronger than Deep Blue now run free on an ordinary phone, and nobody calls them AI. They are chess engines. Finding a fast route through a road network was a research problem for AI labs in the 1960s; it is now called directions. Predicting your next word was a research frontier for decades. On a phone keyboard it is called autocomplete.

The pattern only runs one way. While software struggles with a task, the task belongs to artificial intelligence. Once software does it reliably, cheaply and everywhere, the task gets renamed after the job. AI is the name software carries while it still surprises us; afterwards it is called a spam filter, a chess engine, or directions. Artificial general intelligence, or AGI, is the name for the destination that motion is measured against: a system general across the range of things people do, rather than one that is superhuman at a single task and then renamed after the job. In a headline, AI is a claim about how a piece of software was built, and AGI is a claim about how much of what people do one system could cover.

So the map above has a time axis. Some of what is loudly AI today will keep the name; some will fade into the furniture and lose it, the way chess did. The label tracks novelty at least as much as it tracks technology, and that is worth remembering whenever a headline announces that AI has arrived in some industry. Often the trained systems arrived years earlier, quietly. What arrived this year is the kind that talks.

What the label is worth on a box

Because the label sells, it gets stuck where nothing was learned. A thermostat that follows a fixed schedule can ship with AI printed on the box, and often there is nothing underneath to earn it. The check is to ask the question the label stands in for: what did this learn, and from what examples? A trained system has an answer, even when the company keeps the details vague. The spam filter learned from mail people flagged. The photo search learned from pictures paired with descriptions. A schedule with a sticker on it has no answer at all.

Keep that question handy for a second reason: people have always been generous with this label. The first chatbot, ELIZA, was built in 1966 from a short stack of hand-written text-shuffling rules, and its creator published it partly to show how little was underneath. Users confided in it anyway. Some of what makes software seem intelligent has always been supplied by the person looking at it. The label is applied by marketing from one side and by our own impressions from the other, which is exactly why it cannot be trusted to do your sorting for you.

"Is this AI?" turns out to have no stable answer. The chess engines show that, and so does ELIZA. "What did it learn, and from what?" has an answer every time, and the whole subject gets easier the moment you start asking it.