Moravec's Paradox: Where AI Actually Pays Off

Ask most business leaders where AI will hit first and you get a predictable answer: the low-skill, repetitive, "anyone could do it" jobs. It feels obvious. It is also, in a strict technical sense, backwards. The tasks a human finds effortless are frequently the ones AI struggles with most, and the tasks that take people years of training are often the ones AI handles in seconds. This is Moravec's Paradox, and once you see it, you stop guessing about where AI pays off and start knowing.

What Moravec's Paradox actually says

In the 1980s, the roboticist Hans Moravec and his contemporaries noticed something that broke their intuitions. It was comparatively easy to make a computer perform at adult level on intelligence tests or play strong chess — the things we treat as the pinnacle of human intellect. It was fiendishly hard to give a machine the perception and mobility of a one-year-old: recognising a face, walking across a cluttered room, picking up a mug without crushing it. As Moravec put it, the hard problems are easy and the easy problems are hard. The skills we are proudest of turned out to be the cheap ones to automate; the skills we barely notice ourselves using turned out to be the expensive ones.

Why the easy things are hard

The explanation is evolutionary. Sensory-motor skills — seeing, moving, grasping, reading a room, sensing when a customer is quietly unhappy — have been refined over hundreds of millions of years. They run below conscious awareness precisely because they are so heavily optimised; you don't feel yourself doing them, so they seem simple. Abstract reasoning, algebra, formal logic, structured writing — these are evolutionary newcomers, a few thousand years old at most. We find them hard and effortful, so we assume they represent "real" intelligence. But because they are recent and deliberate, they are also explicit, rule-bound and data-rich — exactly the conditions under which machines excel. The difficulty you feel is not a measure of how hard a task is to automate. Often it is the opposite.

What this means for AI in your business

Modern AI has widened the paradox rather than resolved it. Large language models are, in effect, engines for the abstract layer — pattern, language, structure, inference over text and numbers. That is why they are startlingly good at the parts of knowledge work that once signalled expertise: reading a fifty-page contract and surfacing the awkward clauses, drafting a first version of a proposal, reconciling messy spreadsheets, summarising a research field, generating a forecast from historical data, translating between languages, writing and debugging code. If a task is mostly "take this information and produce that information," it sits squarely on AI's strong side.

The weak side is everything that leans on a body or on genuine social judgement. A plumber diagnosing a leak in an awkward crawl space, a nurse noticing that a patient "just doesn't look right," an electrician improvising around a wall that isn't where the drawings said, a salesperson sensing the moment to stop talking, a manager delivering hard news with the right amount of warmth — these blend fine motor control, live perception and social read-outs that no dataset fully captures. They look like the "simple," low-status parts of a job. They are the parts AI is worst at and, for now, most likely to leave untouched.

A simple test for picking AI use cases

You can turn the paradox into a screening question for any candidate use case: does this task live mostly in information, or mostly in the physical and social world? If the inputs and outputs are documents, data, text and code, it is a strong candidate — automate or augment it, and expect real leverage. If it depends on hands in the field, bodies in a room, or reading unspoken human signals, be sceptical of a full-automation promise and think instead about assisting the person who does it. A useful sharpening: the closer a task is to something a skilled professional does at a desk, the more AI-ready it usually is; the closer it is to something a toddler does effortlessly — grasping, moving, recognising, empathising — the harder it stays.

The practical consequence is that the highest-return AI projects are rarely the ones that replace a whole role. They tend to carve the information-heavy work out of a job — the reading, drafting, checking, summarising and forecasting — and hand the physical and human parts back to the professional, who now spends more time on exactly the judgement AI can't do. Getting that split right, rather than chasing a machine that does everything, is where most of the value sits. If you'd like a second pair of eyes on which of your processes fall on which side of the line, that is the kind of question our team works through in AI consulting engagements — but the paradox alone will already tell you most of what you need to know.

Our AI Act Check (€950 fixed) gives you your risk classification, obligations and a concrete to-do list within one week — including what you can safely ignore.

Frequently asked questions