External signal·Benedict Evans·May 24, 2026·6 min read
Predicting AI job exposure
“The job is a complex mesh of things that we lack the capability to explain explicitly.”
Summary
Benedict Evans argues that trying to predict which jobs AI will hit by scoring "exposure" occupation-by-occupation is fundamentally unreliable. His central evidence is historical: a century of automation in accounting — including the spreadsheet, which let one person do in days what once took a team — coincided with more accountants, not fewer. Jobs keep their titles while their actual content quietly transforms, and technology often disrupts not the task but the business model that funds it, as the internet did to journalism. He points to second-order shocks no exposure model would have caught — smartphones collapsing taxi-medallion values via Uber — and to the crudeness of task databases like O*NET. The deeper problem, he says, is that we can't forecast the future of jobs because we don't even have an explicit, complete description of what today's jobs actually are.
Predictions for the future of work
Evans deliberately resists point predictions — his thesis is that confident, granular forecasts of AI job losses are a category error. He expects AI to reshape work substantially, but in ways that are unknowable in advance: roles will be redefined rather than cleanly deleted, new jobs and whole business models will appear, and the largest effects may come from indirect cascades rather than direct task automation. The practical takeaway is humility — treat "exposure" scores as directional curiosities, not roadmaps, and expect the exceptions to outnumber the rule.
Originally published by Benedict Evans · May 24, 2026
Read the original at Benedict Evans