The reading room · Digna Legi
Predicting AI job exposure
/100
80–100: high value. 70–79: worth the time. Below 70: below the usual publication threshold.
Evidence-reviewed score based on available publisher text. The argument is strong as model criticism, but it does not provide a replacement forecasting method.
Scores reflect one reader’s profile, not an objective quality rating. Best is a separate personal selection.
How scoring works →This brief · about 3 min with detail
Why read this
Evans argues AI job-exposure scores invite false confidence when technology changes demand, job content, categories, and surrounding business models.
AI brief · Checked against source text
The main idea
Evans argues that quantified AI job-exposure models are weaker than they look because they treat jobs as stable bundles of tasks. Past technologies did not just replace tasks; they made activities cheaper, expanded demand, changed job content under old titles, created and erased categories, and disrupted business models around apparently unaffected work. His claim is not that nothing can be predicted, but that job-by-job precision becomes false confidence early in a broad technological shift.
Go a little deeper
Cheap work can create more work
The accounting example separates substitution from demand expansion. If a discounted cash-flow analysis falls from a week to seconds, firms may not simply employ fewer people to do the old volume; they may run more analyses, use them in new decisions, or create different finance work around the cheaper capability. Exposure, in this account, can be a growth channel as well as a threat.
Job titles hide moving targets
Evans stresses that labor categories are not stable measurement units. The same business purpose can move through different titles as tools change, while an unchanged title can cover substantially different work. That makes longitudinal exposure scoring treacherous: a chart may look as if it is tracking one occupation when the lived job, surrounding tools, and classification labels have shifted underneath it.
The business can break while the job survives
Some technology shocks leave visible craft requirements mostly intact but destroy the economic system paying for them. Journalism and music are used to show how internet distribution attacked classifieds, manufacturing, shipping, and local monopoly structures rather than merely the skill of editing copy or finding artists. AI exposure may similarly arrive through a company’s dependencies, not through the worker’s own task list.
Task descriptions miss tacit complexity
The O*NET critique is not just that taxonomies are messy. Evans argues that many jobs are too subtle and context-dependent to be fully described as explicit steps, much like failed expert systems tried to reduce recognition and translation to logical rules. Visible outputs such as slides, legal drafts, or code are not necessarily what customers are buying, so output-based automation demos can exaggerate job replacement.
A case from the article
Uber as a failed back-test
Evans says mobile specialists in the 2000s talked extensively about location data, yet did not identify taxis as a field whose work and economics could be transformed. More efficient dispatch might have seemed plausible; the destruction of medallion value and a changed job structure did not. The example illustrates his test for AI forecasts: would the method have caught indirect, platform-level disruption before it was obvious?
How the case is made
The case is made through historical back-tests against accounting automation, media distribution, smartphones and taxis, plus critique of occupational taxonomies.
Where the idea has limits
The argument still allows directional claims, especially about repetitive clerical work; its target is quantified occupation-by-occupation certainty, not every broad exposure heuristic.
A question to take away · from Digna Legi
Which apparently peripheral business dependency, not which task, would most embarrass your forecast if AI made it cheap?
What the original adds
The source adds detailed historical texture: CPA demand after VisiCalc-style productivity gains, census categories that appear and disappear, and a proposed newspaper/Uber/CPA back-test for exposure models.
About this brief
AI-written, then separately checked for source support, useful detail and clarity. The author’s claims and our editorial question are kept separate. The original remains the author’s work. How we select and summarise →
How was this brief?
Rate this summary, separately from the author’s article.
Optional. Saved in this browser; shared only if you allow analytics.
How was the original article?
Rate the author’s original after reading it.
Optional. Saved in this browser; shared only if you allow analytics.
Digna legi. Worth reading.