Where AI already does real work

Between your card touching the reader and the terminal saying yes, something decides whether the payment is really yours. It takes about a second and no person is involved. Software that learned from an enormous record of past payments, and how each turned out, scores this one, and the score settles it: approve, decline, or ask.

That software is wrong in both directions every day. It declines people buying their own groceries and it waves through some theft. It has been in service anyway, for years, almost everywhere. What keeps it there is not that its error rate is small. It is that the bank knows what to do about the errors: a text message asking whether that was you, a queue where the expensive cases go to a person.

Look anywhere AI does real work and you find that arrangement. What differs is the answer to one question: what happens when it is wrong?

Where the mistakes go

Ranking got in early, and it cheated on that question to do it. Nobody can say which post should have been third. There is no right answer to be wrong about, only a guess that holds attention slightly better or worse than the last one. Software with no wrong answers available to it can go into service the moment it works.

Most work is not so forgiving, and the next cheapest arrangement is an undo. Your camera runs on one. The picture you took was not really taken, it was computed: several frames merged and sharpened by software trained on what photographs are supposed to look like. It survives being wrong because a failure costs one bad picture and a retake. Nobody compares a photograph against the light that fell on the sensor.

Where a mistake is expensive, the arrangement gets expensive too, and what it buys is usually a second pass. Machine translation moves an enormous volume of text every day, and the translations that carry weight go to a person who edits rather than writes. Transcription works the same way: the subtitles, the doctor's dictation. In none of these is the machine's job to be right. Its job is to be close enough that checking it is cheaper than doing it.

An earlier page put the spam filter and the feed on your map of an ordinary day. This is the other half of that map: the same kind of software working on your behalf in places you have no reason to look, and the reason you have no reason is the point. The AI you never notice is not the AI that stopped making mistakes. It is the AI whose mistakes stopped being yours.

The ones that hand you the mistake

Every system above judges, sorts or converts, and what comes out is small: a score, an order, a line of text, a picture nobody will audit. The ones that arrived loudly make things instead. A chatbot's answer, a generated image, a suggested block of code, a voice reading a paragraph aloud. Each is a finished object made for one person on request, with no queue behind it and no second pass. You are the second pass. That is why these landed as products you sit in front of, not machinery humming somewhere.

Code is the case that shows the rule working. A programming assistant is wrong constantly, and it went into daily professional use fast, because programming already had the checking built. The code compiles or it does not. The tests pass or they do not. Somebody reads it before it ships. A craft that spent decades building machinery to catch human mistakes turned out to own machinery that catches a machine's mistakes just as well. Where that kind of checking is thin, the uptake is thin with it.

Image generation sits at the same end for a cruder reason. In August 2022 the trained settings of an image generator were published for anyone to download, and within a week strangers had built things with it nobody had planned. A bad picture costs one more press of the button, the cheapest second pass there is.

A prediction somebody could check

The strongest thing that can catch a mistake is an experiment, which is why the clearest case of AI doing scientific work happened in a field that already knew how to find out.

Every two years since 1994, protein scientists have run a blind exam. Groups are handed nothing but the chain of chemical units a protein is built from, and asked what shape it folds into. Their answers are scored against a shape some laboratory has already established and not yet published. In November 2020 a trained system entered that exam, and its organisers announced that a fifty-year-old problem in biology had been solved.

What made the predictions usable was not that they were always right. A predicted shape is a hypothesis, and biology had spent a century building ways to test one. They went into the ordinary work of laboratories that could check them. That is what real scientific use looks like, and it looks very little like a machine making discoveries.

What a launch video leaves out

A demo is the software. A deployment is the software plus everything built to survive it being wrong.

A launch video shows the system succeeding, which is compatible with any failure rate at all. It does not show who is standing behind the system: the operator watching remotely, the queue, the older and duller thing that quietly takes over. And it does not show the conditions. Demos run on chosen inputs, in good light, on the route somebody drove twenty times last week.

You have met this gap from the other side. Every customer-service chatbot that cannot find your order was once a transcript in an announcement, handling a refund beautifully. Nothing was faked. The transcript was real, the thing in front of you is real, and between them lie all the customers who asked in ways nobody rehearsed.

The gap runs the other way too. A deployment is usually duller than the demo that preceded it and worth far more. The score at the card reader was never going to be a launch event.

So there is one question, and it is not the headline's. The headline asks whether the software can do the thing. The useful question is who catches it when it cannot. Where something real catches it — a test that fails, an experiment that settles it, an undo that costs a second — the work can be deployed, and probably has been for longer than you noticed. Where nothing catches it, you are looking at a demo, however well it performs. And where the thing catching it is a person who did not choose to be standing there, the question has stopped being an engineering question.