Expertise

Autor & Thompson — Expertise

TL;DR

An NBER working paper (No. 33941, June 2025; delivered as the Joseph R. Schumpeter Lecture to the European Economic Association, Rotterdam, 29 August 2024, under the title “Does Automation Replace Experts or Complement Expertise? The Answer is Yes.”) by David Autor (MIT Economics) and Neil Thompson (MIT FutureTech / CSAIL / Initiative on the Digital Economy). It proposes and empirically tests an expertise framework that explains a puzzle the canonical automation literature cannot: why routine-task automation has simultaneously lowered employment but raised wages in some routine-task-intensive occupations, while doing the reverse in others.

The two theoretical pillars:

1. Expertise — a worker’s capability to perform specific tasks, rankable along a single hierarchical dimension (following Garicano 2000; Garicano & Rossi-Hansberg 2006): a higher-expertise worker can always perform a lower-expertise task, never the reverse. Expertise is a wage premium and a barrier to entry.
2. Occupational task bundling — all tasks bundled into an occupation must be performed by every worker in that occupation, and expertise requirements differ across the bundled tasks. Some are the occupation’s expert tasks; others are its inexpert tasks. (This departs from canonical task models, which assume atomistic task assignment with no two skill groups performing the same task.)

The core prediction, worked through the paper’s canonical contrast (accounting clerks vs. inventory clerks — both extensively computerized, both facing similar automation exposure): automation that removes an occupation’s inexpert tasks raises the expertise required for what remains → wages rise, employment falls (fewer workers qualify). Automation that removes an occupation’s expert tasks lowers the expertise required for what remains → wages fall, employment rises (more workers can now qualify). Wage and employment effects move in opposite directions, and the direction itself is set by which tasks — expert or inexpert — are automated, not by how much automation occurs.

The taxi-driver / proofreader worked pair (the TEDx talk’s central illustration, §below): GPS automated taxi drivers’ single most expert task — memorizing streets (London’s famous multi-year “Knowledge” test) — so driving became accessible to far more people: wages grew more slowly than the economy; employment surged (the Uber-era expansion). Spellcheck automated proofreaders’ least expert task, leaving the harder judgment work (argument structure, evidence quality) intact: wages grew significantly faster than the economy; employment shrank.

The empirical contribution: a novel, content-agnostic method for measuring occupational task expertise from job-task text, grounded in the Efficient Coding Hypothesis (word frequency/entropy), plus a longitudinal method using word embeddings to track which tasks were added or removed from occupations across four decades — without requiring tasks to be described consistently over time. Applied to U.S. occupational employment and wage data (1980s–2020s):

• Changes in occupational expertise (from task removal and addition) strongly predict changes in occupational wages, independent of the sheer quantity of tasks added or removed.
• Removing expert tasks and adding inexpert tasks both predict wage declines; removing inexpert tasks and adding expert tasks both predict wage gains — the reverse of quantity effects (gaining tasks expands employment; losing tasks contracts it — opposite to the expertise-wage relationship).
• Employment moves opposite to wages on the expertise axis: occupations with rising expertise requirements see falling employment (fewer qualify) despite rising wages; occupations with falling expertise requirements see rising employment despite falling wages.
• Applied to the historical routine-task-automation literature specifically: automation bifurcated occupational outcomes depending on whether the automated routine tasks were relatively expert (→ lower wages, higher employment in that occupation) or relatively inexpert (→ higher wages, lower employment) — resolving the “why did routine automation lower employment but often raise wages” puzzle that canonical task models leave unexplained.
• The authors explicitly flag the AI era as a next natural application of the same framework (Conclusions, p.53): “there is nothing in our model that is specific to the computer era… the rapidly advancing era of Artificial Intelligence” is named as an equally applicable domain, though the empirical evidence in the paper itself covers the last four decades of computerization, not AI specifically.

What was actually ingested

The full 74-page working paper — abstract, introduction, theoretical model, the novel expertise-measurement methodology, the four-decade empirical analysis, the routine-task-automation application, conclusions, references, and appendix (including the GPT-4.1 prompt used to classify tasks into the RC/RM/NC/NM/NI routine-vs-non-routine × cognitive-vs-manual taxonomy). Converted from the co-located PDF (raw/papers/expertise-autor-thompson-2025.pdf, gitignored) via pdftotext -layout; page count (74) and the presence of a complete References section and Appendix confirm no truncation.

Why this source matters to the wiki

This is the wiki’s first primary academic paper formalizing occupational task bundling with an expertise hierarchy — a theoretical grounding the wiki previously held only at the level of practitioner restatement (Evans’s task vs job) or a different formal tradition (Jones’s macro weak-links growth model). It supplies:

• The mechanism behind the wiki’s “automation sometimes raises wages” puzzle. ai-employment-effects already documents that automation-exposed occupations show heterogeneous wage/employment outcomes; this paper is the theoretical account of why, keyed to which tasks (expert vs. inexpert) get automated — not merely how much of an occupation is automated.
• A rigorous companion to Jones’s weak-links model. Both predict that partial automation raises the wages of workers retained in the un-automated remainder; Jones’s version is macro/growth-theoretic (jobs as generic task bundles, GDP-share accounting), Autor-Thompson’s is micro/occupational (a rankable expertise hierarchy, tested against 40 years of U.S. occupational payroll and task data). The taxi-driver/proofreader pair is a sharper, more falsifiable worked example than Jones’s radiologist anecdote.
• A theoretical account for Brynjolfsson’s codified-vs-tacit mechanism. Brynjolfsson et al. find AI displaces codified knowledge more than tacit knowledge, disproportionately hurting young workers who supply more of the former. Autor-Thompson’s expert/inexpert task distinction is a formalization of the same codified/tacit intuition, with an explicit hierarchy and task-bundling structure that makes it empirically testable at the occupation level.
• An accompanying TEDx talk (Thompson, TEDxBoston, July 2026) that translates the same expertise framework — using the identical taxi-driver/proofreader pair — into an AI-era register, plus a new empirical layer (task-level LLM evaluation data) that projects expertise change forward rather than only measuring it retrospectively over the last four decades.

Linked entities and concepts

• ai-employment-effects — primary target; adds the expertise-hierarchy/task-bundling theoretical model as the mechanism behind the page’s heterogeneous automation wage/employment findings.
• automation-vs-augmentation — the expert/inexpert task distinction sharpens the task vs job analytical lever already on that page (Evans) with a formal, empirically-tested hierarchy.
• Dangling (single-source mention, deferred per Author-entity promotion): David Autor (MIT Economics; co-author) — named on this source only so far; promote on a second-source mention.

Source quality

Peer status: an NBER working paper (circulated for discussion; not yet peer-reviewed, per NBER’s standard disclaimer) delivered as an EEA Schumpeter Lecture — the format the discipline uses to showcase a senior economist’s (here, joint) synthesis-and-new-result contribution. David Autor is one of the most-cited labor economists working on automation and task models (co-author of the canonical Autor-Levy-Murnane 2003 task model and the Acemoglu-Autor 2011 task-model literature this paper explicitly revises); Neil Thompson directs MIT FutureTech. Treat the theoretical model as rigorous and the empirical results as illustrative rather than definitive per the authors’ own conclusion (p.53): a “relatively coarse set of occupations studied over a relatively short interval.” No funding conflicts of the kind that would bias the automation-wage findings (funders: Hewlett Foundation, Google Technology and Society Visiting Fellows, NOMIS Foundation, Schmidt Sciences AI2050, Smith Richardson Foundation, Open Philanthropy).