Predicting AI job exposure — Benedict Evans
Benedict Evans · Benedict Evans · 2026-05-24
Trying to rank jobs, companies, or industries by their exposure to AI looks rigorous but is really an attempt to forecast second- and third-order effects that even domain experts have always missed.
The hard part of a new technology is never the obvious automation of a visible task — it's the business models that quietly collapse and the new ones that appear from directions no expert was watching. A century of accounting automation grew the profession, the internet hollowed out newspapers through classifieds rather than writing, and smartphones killed taxi medallions despite mobile insiders thinking only about better dispatch. Any AI exposure methodology that wouldn't have caught those three should be treated as decoration, not forecasting.
Attempts to score jobs, companies, and industries by their exposure to AI are exercises in predicting something that cannot be predicted. The viral charts and tables built from census data give an illusion of rigor without real forecasting power.
A second class of disruption leaves the actual work unchanged but destroys the business model that paid for it. The journalist and the A&R scout were unaffected by the internet directly — their employers' revenue streams were not.
Calculating machines, mainframes, databases, spreadsheets, ERPs, and cloud were all built largely to automate accounting, yet the number of accountants kept rising. CPAs are the profession that any 1970s automation-exposure model would have put at the top of the kill list.
People deep in mobile and location data in the 2000s never flagged taxi drivers as exposed to smartphones — at most they imagined better dispatch. The biggest second-order effects of a new technology are invisible to the experts closest to it.
A minimum sanity check: would your exposure model have caught that automation grew accounting, that the internet broke newspapers via classifieds, and that smartphones destroyed taxi medallions? If not, it offers little to the rest of us.
Open
- · If exposure scoring doesn't work, what kind of analysis would actually help workers, firms, or policymakers prepare for AI?
- · How do you distinguish a useful structural observation about AI from a confident-sounding prediction that will age badly?
Pipeline
- source kind
- url
- generated by
- anthropic+voyage
- candidates
- 20 (selected 5)
- embeddings
- voyage-3.5
Coverage
100% covered
Each block is one paragraph of the source. Darker means the decomposition captures it well; lighter means it was left out — the part of the document the summary doesn’t cover.
Considered candidates (15)
Below top-k · 10
- claimJobs cannot be usefully described in the first placec 0.90
Even setting change aside, you can't write down what a job actually is in enough detail to reason about its automation. The work is a complex mesh of tacit components that resists explicit description.
- evidenceBack-testing past tech shifts inverts the predictionsc 0.80
When you apply the same methodology to previous technology waves, industries that should have been devastated often grew, and industries that did get hit should by the logic have been immune. The track record of this kind of analysis is bad.
- mechanismJob descriptions repeat the failure of expert systemsc 0.80
Trying to enumerate what an associate at a law firm does fails for the same reason hand-coded AI failed at image recognition — the underlying activity is too subtle to specify. Real jobs are rarely a single task that compresses into a button.
- mechanismGell-Mann amnesia about AI's reachc 0.80
You know how irreducibly complex your own field is and how partial an AI demo really is, but in other fields you see a slick template and assume the profession is doomed. When you hire McKinsey or buy software, the deliverable is not the product.
- mechanismJevons paradox turns exposure into more workc 0.75
When automation makes a task cheap, the response is often to do far more of it rather than do the same amount with fewer people. If a DCF goes from a week to thirty seconds, you do many more DCFs — exposure to automation can mean more work, not less.
- caveatO*NET-style task lists miss the real shape of jobsc 0.75
Frameworks that decompose a job into a list of automatable tasks tell you nothing about how the job grows, shrinks, or transforms with automation, nor about disruption happening outside the analysis. The methodology is structurally blind to the things that actually matter.
- caveat'Directionally correct' is not a useful answerc 0.75
Saying repetitive clerical work is most exposed is true on average but useless in practice, because the exceptions may dwarf the rule. 'Physical media distribution will be disrupted' was directionally correct in 1995 yet meant wildly different things for newspapers, labels, TV, and film.
- evidenceJob titles drift while underlying work stays the samec 0.55
Census categories like 'billing machine operator' or 'data keyer' appear and disappear, often capturing the same person doing the same business purpose under a new label. Title-based analyses confuse taxonomic churn with real change.
- contextWhy 'it depends' is the honest answerc 0.50
Critics dismiss hedged conclusions, but at this stage of a new technology, specific predictions are right only by luck. The intellectually serious move is to name the framings rather than fake the precision.
- contextNon-technology variables also move the numbersc 0.40
Regulatory changes created one-off surges in CPA hiring that have nothing to do with automation. Any exposure analysis ignores the ceteris paribus problem at its peril.
Redundant with selected · 5
- mechanismAutomation unlocks new kinds of work, not just more of the oldc 0.85 · sim 0.83
Making something cheap and quick usually unlocks adjacent activity, so the work transforms rather than scales. Accountants today aren't doing 1970s accounting 'but more' — the job title is stable while the job itself has changed.
overlapped with: A century of automating accounting only grew the profession
- implicationFrameworks beat quantified forecasts at this stagec 0.85 · sim 0.84
At the start of a fundamentally new technology, mental models and historical analogies are honest; specific job-by-job radar charts are self-deception. You don't know what the jobs are today and you don't know how they'll change.
overlapped with: Predicting AI job exposure is fundamentally impossible
- implicationAI will indirectly disrupt jobs it can't doc 0.80 · sim 0.83
Plenty of people will have AI-resistant jobs sitting inside businesses whose defensibility relies on something AI eats. The exposure that matters may be one or two steps removed from the work itself.
overlapped with: Predicting AI job exposure is fundamentally impossible
- exampleNewspapers and record labels were broken by their funding stackc 0.70 · sim 0.86
Journalism was subsidized by a local monopoly on classified ads; record executives were paid by manufacturing plastic discs. The internet didn't change the craft, it dissolved the unrelated business that funded it.
overlapped with: The job can be untouched while the business collapses beneath it
- implicationHalf the affected jobs may be ones you didn't listc 0.65 · sim 0.84
An exposure analysis can be confidently wrong in both directions: jobs flagged as doomed survive untouched while huge pools of unflagged jobs get transformed. The error bars on this kind of forecast swallow the forecast.
overlapped with: Predicting AI job exposure is fundamentally impossible
Janitor
Non-content spans (acknowledgements, references, footnotes, headers, boilerplate) are dropped before the decomposition runs.
- total spans
- 19
- kept
- 18
- dropped
- 1
- content · 18
- noise · 1