AI and Work

Here's the counterintuitive part. The tasks most exposed to automation are rarely the ones we think of as low-skill. For decades, the safest assumption was that machines would take the manual, repetitive work first and leave the thinking to us. Generative AI has scrambled that order. The systems that can now draft a competent legal memo, write serviceable code, or summarize a research paper struggle to reliably fold laundry or fix a leaking pipe under a sink. The cognitive, credentialed work has turned out to be more exposed than the physical, improvised kind.

This has a name that predates the current wave: Moravec's paradox, after the roboticist Hans Moravec, who noticed in the 1980s that the things evolution spent millions of years perfecting — walking, grasping, perceiving a cluttered room — are the hardest to reproduce in a machine, while the things humans find effortful, like arithmetic or formal logic, come easy to a computer. What feels advanced to us is often computationally cheap. What feels trivial is often the deepest problem in the field.

If tasks are the unit, then the honest way to read the labor market is to watch how they get recombined. When part of a role gets automated, the remaining tasks don't just sit there. They get rebundled into a new version of the job, or split off into a different one, or absorbed into someone else's day. The occupation label may stay the same while the work underneath it changes almost completely — which is why the raw employment count can look stable even as the ground shifts.

Consider what the current wave is doing to writing-heavy work. A marketing team that once needed five people to produce a steady volume of copy might now need three, because a model handles the first draft and the humans edit, direct, and decide. The occupation "copywriter" doesn't vanish from the statistics. But its center of gravity moves from generating text to judging it, and the number of people the work supports quietly contracts. That's not a robot apocalypse. It's a slow reweighting, and it's much harder to protest because no single day is the day the jobs died.

Step back from the ledger of jobs saved and jobs lost, and a more useful picture appears: automation is constantly redrawing the map of which human skills are scarce and which are cheap. Every task a machine absorbs pushes value toward the tasks it can't. When ATMs took the cash-handling, the value of a teller's warmth and sales judgment went up. When models take the first draft, the value of taste, direction, and accountability goes up. Automation doesn't so much destroy value as move it — and the people who thrive are the ones standing where it moves to.

This reframes what "exposure" really means. A job isn't safe or doomed; it's a portfolio of skills, some of which are appreciating and some depreciating at any given moment. The unsettling feature of the current wave is how fast it revalues skills that felt permanent. Fluent writing, competent coding, basic analysis — capabilities that took years to build and reliably commanded a salary — are becoming abundant. Abundance is good for whoever buys the skill and hard for whoever sold it. That's the quiet drama inside every calm employment chart.

Frey and Osborne's number wasn't a lie; it was an answer to the wrong question. They measured which jobs contained automatable tasks and reported it as if the jobs themselves were on the block. A decade of stable employment and quietly transformed workdays showed the difference. Machines came for tasks, the tasks got rebundled, and the occupation labels mostly held while the work inside them turned over. The apocalypse arrived, but disguised as a reweighting nobody could point to on a calendar.