Moravec's paradox: why AI can pass the exam but not fold the towels

In 1988 the roboticist Hans Moravec wrote down something that everyone building robots had quietly noticed. Getting a computer to play draughts or pass an intelligence test was comparatively easy. Getting it to see and move as well as a one-year-old was difficult or impossible. The hard things were easy. The easy things were hard.

It’s now called Moravec’s paradox, and if you’ve ever wondered why the scores on this site look the way they do, it is most of the answer.

Why the easy things are hard

The usual explanation goes like this. Walking, grabbing, spotting a face in a crowd: evolution has been polishing those for hundreds of millions of years, so they feel effortless to us, and we massively underestimate how much is going on. Algebra, chess and filling in forms are brand new by comparison. They feel hard because we are bad at them, not because they are deep.

So when engineers started building thinking machines, they started with the stuff that felt hard, and were delighted by how quickly it fell. Then they tried to make a robot pick up a sock.

What it looks like in the data

Look at the two ends of the league table and you can see Moravec everywhere.

At the bottom are jobs made mostly of the “hard” things: reading, sorting, typing, calculating. Data entry clerks score 30, the lowest on the site. Every one of their tasks is the kind of thing a computer found easy decades ago, and a language model finds easier still.

At the top are jobs made of the “easy” things. Kitchen helpers score 87. Domestic cleaners score 86. Plumbers score 85. Nobody puts “scrubbing a pan” on a list of great intellectual achievements, but it takes eyes, hands, balance and a feel for how hard to press, all at once, in a kitchen that’s different every time. That’s Moravec’s list.

The third figure on every job page (“how fiddly and cramped”) is Moravec’s paradox turned into a number. It comes from Frey and Osborne, who picked finger dexterity, manual dexterity and working in cramped, awkward spaces as the bottleneck that had held machines back. They were measuring exactly the “easy” things that turned out to be hard.

The big asterisk

Here’s the bit a good-news story would skip.

The paradox is about robots, and the main score here is about AI software. A job scoring 85 against a language model is not a job a robot has failed at. It’s a job a language model can’t reach, because the work isn’t made of words. Robots are a separate question, and a lot of money is going into answering it. That’s why every hands-on job on this site carries a warning that the score is probably too hopeful.

Some people think the paradox is partly an illusion. The computer scientist Arvind Narayanan has argued that it’s partly a selection effect: researchers work on the problems that look promising, so we mostly hear about the “hard” things that turned out easy, and rarely about the ones that stayed hard. It’s a fair challenge, and a good reason to treat the paradox as a pattern that has held so far rather than a law.

And “hard for machines” is not “safe forever”. Robot vacuums already do a chunk of one of the “easy” jobs. Warehouses are full of machines doing picking that was meant to be impossible. The paradox tells you which way round the difficulty has run for forty years. It doesn’t tell you it will run that way for another forty.

What to take from it

If your job is mostly words, numbers and screens, the paradox isn’t on your side. That’s not a reason to panic; it’s a reason to look at your own tasks and notice which of them need judgement, trust or a body in the room.

If your job is mostly hands, balance and awkward spaces, you have Moravec on your side for now, and a lot of engineers trying to prove him wrong. Keep an eye on the robots, not the chatbots.

And if you ever feel daft for finding the spreadsheet easy and the fitted sheet impossible: that’s not you. That’s hundreds of millions of years of evolution, and the robots are finding the fitted sheet impossible too.