Skip to content
Main Site News Console

Fact-Checking the Moravec Paradox

· Normal Tech (AI Snake Oil) Translated
分析实践

I run a YouTube channel where I analyze the evolution of AI from a general technical perspective. This article is based on my recent video, in which I take a deep dive into Moravec’s paradox—the oft-repeated maxim that things humans find hard are easy for AI, and vice versa.

Here are my findings:

  • Moravec’s paradox has never been empirically tested. (Many AI researchers, including some pioneers I know and respect, often repeat it as fact—but that doesn’t mean I’ll take it at face value!)

  • What it really describes is what the AI community thinks is worth spending effort on; it does not predict which problems will be easy or hard for AI.

  • It also comes bundled with what I consider a highly questionable evolutionary explanation. (AI researchers have always loved making up claims about the human brain without any background in neuroscience or evolutionary biology.)

  • Moravec-paradox-style thinking has led both to alarmism (about imminent superintelligent reasoning) and to false reassurance (in robotics and similar fields).

  • To adapt to AI progress, we do not need breakthrough prediction powers. Because it takes a long time for new capabilities to spread, we have plenty of time to react—and we usually waste that time, then panic!

You can watch the full argument here, or keep reading below.


Every week, we hear new claims about progress in AI. How do we know what will happen next? Can AI predict crime? Write award-winning novels? Hack critical infrastructure? Will we finally get robots that can help us fold laundry and load the dishwasher at home?

What will AI progress mean for your job? For the social fabric? It’s genuinely hard to navigate all this uncertainty. If only there were some way to predict which new AI capabilities will be developed soon, and which ones will remain out of reach for the foreseeable future.

Historically, AI researchers have beenpretty bad at forecasting the progress of AI capabilities. We really don’t have any principles that tell us which tasks will be easy for AI and which will be hard.

But we do have one: Moravec’s paradox. It refers to the observation that it is easy to get computers to do things humans find difficult, like math and logic, but hard to get them to do things humans find easy, like seeing the world or walking around in it.

The idea comes from Hans Moravec’s 1988 book Mind Children. Moravec was, and is, a robotics researcher. He wrote:

It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.

In the early days of AI, researchers focused on chess and other reasoning tasks because these were thought to be the hardest and most distinctively human. Interestingly, if you want to build a machine that can beat a human grandmaster at chess, figuring out the move is the easy part. The hard part is physically making that move on the board. Today, that is widely understood, and so Moravec’s paradox seems to fit our intuitions pretty well.

If Moravec’s paradox were true, its implications would be enormous. If we want to know what AI capabilities are likely to emerge next, all we would have to do is look at how hard they are for humans. Then science would get automated before folding laundry, and so on.

But here’s the thing: Moravec’s paradox has never been fact-checked. Despite being repeated as fact in videos with hundreds of thousands of views and even in TED talks, that remains true. When I dug into the evidence behind the so-called paradox, I found something surprising.

In this article, I’ll discuss why both the theory and the evidence behind the paradox are shaky. Then I’ll explain why simplistic predictions about what AI will find easy or hard have misled AI researchers and tech leaders alike. They have led to alarmism on the one hand, and false reassurance on the other. (Now that’s a real paradox.) Finally, I’ll answer the question: if we can’t rely on Moravec’s paradox, how should we prepare for AI progress and its effects?

The evidence behind the paradox is shaky

How do we test Moravec’s paradox? We could take a bunch of real-world tasks, estimate how hard they are for humans and how hard they are for AI, and then plot them against each other. If we saw something like this, then the paradox would be confirmed.