No. 111AI Systems

Moravec's Paradox

What is hard for people is often easy for machines and what is effortless for people is often hard, so intuition about which tasks AI will find easy is unreliable.

What feels effortless to people is often hardest for machines, and the reverse, so intuition about what AI can do is unreliable.

To an executive

AI can draft a contract yet stumble on what any clerk knows. Which work is easy for AI is not obvious, so test real tasks before planning around them.

To an engineer

Don't rank tasks by how hard they feel to humans. Prototype on real cases and look for failures in context, common sense and consistency, not only in difficult reasoning.

The idea

In Mind Children (Harvard University Press, 1988), the roboticist Hans Moravec observed a reversal that surprised early AI researchers. It had proved comparatively easy to make computers perform at adult level on intelligence tests or at games like checkers, and very hard to give them the perception and mobility of a one-year-old. The skills people find hardest, such as formal reasoning, were the easiest to mechanise; the skills people never notice using were the hardest.

His explanation was evolutionary. Perception and movement have been refined over hundreds of millions of years and run unconsciously, so they feel easy. Abstract reasoning is recent and effortful for us, so it feels hard, but it needs far less computation.

The paradox holds only partly today. Language models have made much "hard" knowledge work cheap: drafting, summarising, coding, answering exam questions. Physical work, reliable judgement in messy context and consistent common sense remain hard. The lasting lesson is that our sense of which tasks are easy for a machine is a poor guide.

Why it works

Human difficulty is not machine difficulty. We judge a task by how much effort it costs us. Machines have a different profile, so that judgement misleads in both directions.

Hidden skills carry hidden weight. Much real work depends on context people absorb without noticing: who the customer is, what the policy really means, which exception is normal. Those parts fail first when a task is automated.

Testing beats intuition. Because the boundary is irregular and moves with each model generation, only trying real tasks shows where it lies.

Where it applies

Choosing AI use cases, where planners rank processes by how senior or skilled the people doing them are.
Process redesign, to find the unnoticed steps that make a workflow function.
Vendor evaluation, where impressive demos on hard-looking tasks hide failures on mundane ones.
Operations and field work, where physical and situational tasks still resist automation.

How to apply

Distrust the feeling of difficulty. Write down which steps you expect AI to find easy and hard, then treat the list as a hypothesis.
Decompose the task with the people who do it. Surface the tacit checks, context and handoffs that never appear in the process map.
Test on real cases. Run representative work, including exceptions, and record where the AI fails.
Design around the hard parts. Keep people on the context-heavy steps, and automate the parts that proved easy.

Example scenario

A bank planned AI for two processes. Credit memo drafting, done by experienced analysts, was judged too complex; invoice matching in back-office operations was judged trivial. The roadmap put invoice matching first.

The architects ran both on a few hundred real cases. The model drafted credit memos that analysts rated as a good first draft. Invoice matching failed on exactly what clerks handled without thinking: duplicate supplier records across three systems, scanned documents and local conventions about partial deliveries. The roadmap was reversed, and invoice matching waited for a cleaner supplier master.

Questions to ask

Have we tested this task on real cases, or are we guessing that it is easy or hard for AI because it is easy or hard for us?
Which parts of this process rely on judgement, context or physical work that people do without noticing?
Where would an AI that is brilliant at the paperwork still fail at the step a junior does without thinking?

When not to use it

Well-measured tasks, where you already have evaluation results and need no heuristic.
Physical robotics specifics, where engineering evidence matters more than the general observation.

How it misleads you

The paradox can be read as a fixed rule that reasoning is easy and perception is hard, which the last decade has partly overturned. Its value is the warning about intuition, not a prediction of which tasks AI will master. Test each task; do not reason from the slogan.