Arvind Narayanan, writing at Normal Tech (formerly AI Snake Oil), argues that Moravec’s paradox — the claim that tasks hard for humans are easy for AI and vice versa — has never been empirically tested and carries less predictive value than its widespread citation suggests.
What it says
Moravec’s paradox originates from Hans Moravec’s 1988 book Mind Children, in which he wrote that “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.” The observation has since been repeated in academic papers, TED talks, and videos with hundreds of thousands of views, Narayanan notes.
Narayanan argues the paradox fails a basic methodological test. To evaluate it properly, one would need a representative sample of tasks, measure how hard each is for humans and for AI, and check for the predicted correlation. What AI researchers actually do, he contends, is focus on two of four possible quadrants in such an analysis: tasks that are easy for humans but hard for AI (interesting because they are solvable targets), and tasks hard for humans but easy for AI (interesting because they augment productivity). Tasks easy for both, and tasks hard for both, receive little attention and little research investment. The apparent negative correlation, Narayanan writes, follows directly from ignoring two quadrants — not from any property of cognition.
On the evolutionary explanation Moravec offered — that human sensorimotor skill reflects a billion years of evolutionary refinement while reasoning is a recent and immature capability — Narayanan is more direct in his skepticism. He describes AI researchers as having “a history of making stuff up about human brains without any relevant background in neuroscience or evolutionary biology,” and says Moravec’s account belongs in that category. He notes that the 1970s symbolic reasoning systems Moravec praised as “easy for AI” are a poor basis for generalizing about modern large language models.
Narayanan argues the more useful framing is not about predicting which capabilities are next, but about diffusion timelines. Because new AI capabilities typically take a long time to spread through the economy, he contends there is usually more time to respond than panic-driven analysis suggests — though he acknowledges that time is often wasted.
Context
The essay is accompanied by a YouTube video Narayanan published as part of a channel he describes as analyzing AI from “a normal technology perspective.” The piece is the third in a series since AI Snake Oil rebranded as Normal Tech, following the publication of a widely circulated essay titled “AI as Normal Technology.” Narayanan and co-author Sayash Kapoor say they plan to expand that framework into a book, which they expect to complete in late 2026 for publication in 2027.