A Singulariki data story
The 2013 reversal
Scroll to watch every U.S. occupation move from its 2013 automation-risk estimate to its 2025 AI task-overlap exposure — the work once rated safest is now the most exposed.
2013 — Frey-Osborne automation risk
Each dot is an occupation. Colour = 2025 exposure band. Task overlap, not automation.
In 2013, two economists ranked every job by its risk of automation.
Carl Frey and Michael Osborne estimated each occupation's probability of computerisation. Here is every U.S. occupation, placed left (lower risk) to right (higher risk) by that 2013 estimate.
The alarm pointed at routine, manual and clerical work.
Drivers, labourers, cashiers and clerks clustered on the high-risk right. Knowledge and people work — managers, teachers, analysts — sat on the supposedly safe left.
A decade later, we can measure something different.
Not “will this job be automated,” but how much of each job's tasks today's generative AI can assist — task overlap, drawn from millions of measured tasks. Watch the same dots move to that 2025 position.
The reversal: the work rated safest is now the most exposed.
Customer-service reps, teachers, managers and analysts — cognitive, communicative, text-heavy work — travel right. Their tasks overlap heavily with what language models do.
And the work it called doomed is among the least exposed.
Truck drivers, labourers, cooks — physical, in-person, bodily work — travel left. Today's AI assists their tasks the least. (Cashiers and clerks score high on both: routine information work either way.)
Two estimates, a decade apart. A contrast — not a forecast.
Frey-Osborne guessed at whole-job automation in 2013. Task-overlap exposure measures something else entirely in 2025. Neither is an observation of jobs lost, and this is not a time-series. It is two different questions, asked ten years apart, disagreeing in a revealing way.