The New Work Hypothesis
The new work hypothesis is Autor's answer to the recurring fear that technology destroys more jobs than it creates. Using the Census Occupation Index — which has catalogued job titles across eight decades — Autor and collaborators documented that approximately 60% of employment in 2018 was in occupations that either did not exist or were negligible in 1940. New work includes not only obviously technological occupations (software engineer, data scientist) but also entire categories of services (home health aide, yoga instructor, financial planner) that emerged as rising incomes and new technologies created demand for activities that had no prior commercial form. The hypothesis does not deny that automation destroys jobs; it argues that the economy's long-run dynamic is one of continuous task creation at the frontier of what is newly possible or newly valuable.
The Denominator Problem — Contrarian ^ Opus
There is a parallel reading that begins not from the Census categories but from the underlying denominator: total hours of human labor the economy requires. Autor's measurement captures occupational proliferation — new job titles appearing in the index — but is silent on whether the total volume of work is rising or falling. If 60% of jobs are new but the workforce has doubled while productivity per hour has tripled, the real story may be that technology creates specialized niches while reducing aggregate labor demand. The yoga instructor and financial planner exist because rising productivity created surplus income, but their emergence tells us nothing about whether the economy needs more human hours or fewer.
The AI case sharpens this. Previous technologies created new tasks because they were complements — the automobile required drivers, mechanics, planners. AI is increasingly a substitute, performing tasks end-to-end without requiring human supervision at scale. The Trivandrum engineers do the work of a hundred, but Autor's framework asks where the ninety-five go without asking whether the economy will generate ninety-five new roles or ten. The measurement counts occupational categories, not labor hours absorbed. If AI creates five brilliant new job types but each employs a thousand people while displacing ten million, the New Work Hypothesis as stated would record this as confirmation — new categories appeared — while missing the collapse in total demand for human cognitive labor. The denominator, not the proliferation of titles, is the question AI forces into view.
In the AI Story
The hypothesis operationalized an old intuition with new precision. Economists since Schumpeter had spoken of creative destruction, but the New Work measurement allowed quantification of the creative side of the ledger. Autor's collaborator Anna Salomons and others extended the analysis internationally, finding similar patterns across OECD economies: roughly half of current employment consists of occupational categories that did not meaningfully exist three generations ago.
Applied to AI, the hypothesis offers both reassurance and warning. Reassurance: past technological revolutions produced more new work than they destroyed, and there is no a priori reason to assume AI breaks this pattern. Warning: the time scale of new-work creation has historically been slow — decades — while AI's destruction of existing tasks is occurring in years. Whether new occupations can emerge fast enough to absorb displaced workers is the central empirical question, and the answer depends partly on institutional and educational responses that operate on their own slow time scales.
The hypothesis also complicates Segal's narrative in The Orange Pill. Segal describes the Trivandrum engineers as doing work that previously required a hundred people. Autor's framework asks: what happens to the ninety-five? The new work hypothesis predicts that many will eventually find employment in occupations that do not yet exist, but says nothing about the transition period, during which the disruption is concentrated on specific workers whose new work has not yet been invented.
Origin
The hypothesis was developed in Autor's 2024 NBER paper 'New Frontiers: The Evolving Content and Geography of New Work in the 20th Century' with Caroline Chin, Anna Salomons, and Bryan Seegmiller. The paper analyzed Census data across 1940-2018 to classify occupation titles as new or pre-existing, producing the empirical foundation for the hypothesis.
Key Ideas
Most work is new. Approximately 60% of 2018 US employment was in occupational categories that did not meaningfully exist in 1940 — the long-run engine of labor absorption is task creation, not task preservation.
Technology creates and destroys. Every technological wave has destroyed tasks while enabling new tasks; historically the creation side has dominated the destruction side, producing rising employment and wages.
Timing matters. The new-work mechanism operates on decadal time scales while destruction operates on annual ones — the transition gap is where disruption is concentrated and where institutional support is required.
Not all new work is good work. The new occupations created by AI will have their own distribution of wages, autonomy, and meaning, and there is no guarantee they will be better than the ones they replace.
Appears in the Orange Pill Cycle
Measurement Precision, Mechanism Ambiguity — Arbitrator ^ Opus
Autor's measurement is precise and his historical finding is robust: 60% of 2018 job titles were negligible in 1940. The question is what mechanism this measurement reveals. If the mechanism is "technology creates new tasks that require human labor," then AI may not follow the pattern — AI's substitution character differs from prior waves (80% toward the contrarian view). If the mechanism is "rising productivity creates demand for new categories of service," then AI follows it exactly, with the caveat that services may require fewer total hours (60/40 split, depending on how much AI compresses labor intensity within new categories).
The timing point is decisive and Autor names it clearly: creation operates on decades, destruction on years. This is fully confirmed by the historical record (100% Autor's framing). The transition gap is where the entire political economy of displacement plays out, and the New Work Hypothesis offers no prediction about whether institutions can adapt fast enough. That is a separate question, which makes the hypothesis a measurement of the long run, not a forecast of the transition.
The synthetic frame the topic requires is this: new work measures occupational diversification, not labor absorption. Both can be true — categories proliferate while total hours demanded fall — and AI may produce exactly this combination. The hypothesis is correct about what it measures and agnostic about what matters most: whether the economy will need as much human labor in 2050 as it did in 2020, regardless of how many new job titles appear in the index.
Further reading
- Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller. New Frontiers: The Evolving Content and Geography of New Work in the 20th Century. NBER Working Paper, 2024.
- Acemoglu, Daron, and Pascual Restrepo. The Race Between Man and Machine. American Economic Review, 2018.
- Lin, Jeffrey. Technological Adaptation, Cities, and New Work. Review of Economics and Statistics, 2011.