What Counts as New Work?
New work cannot be identified simply by looking at an occupation that is growing. The question here is how researchers turn an emerging specialty into an observable category, and what is lost when that category is treated as if it were the underlying task itself. Once the measurement is clear, readers can interpret prevalence estimates without confusing a careful proxy with a complete census of new work.
The IZA study makes this problem explicit. Workers do not arrive in a national data set carrying a field called “new work.” They report what they do, and those descriptions must be compared with an occupational classification that changes over time. The study’s contribution is to connect those two pieces: detailed descriptions from workers and the Census Bureau’s evolving list of occupational titles.
From a new title to a vintage
The paper defines new work through the Census Alphabetical Index of Industries and Occupations, or CAIO. A title is treated as new when it appears in the index for the first time. The period in which it enters the list is its vintage. A title added in the 2000 edition belongs to the 2000 vintage, while a title added in the 2018 edition belongs to the 2018 vintage.
This definition creates a stable starting point. Instead of asking a researcher to decide informally whether “data scientist” or “conference planner” feels new, it uses the point at which the Census’s occupational index began recognizing the title. That makes comparisons across workers and years possible. It also gives the analysis a time dimension: a new specialty can be compared with older specialties that have had more time to spread.
But a classification is not the same thing as reality. A title can be added after workers have already been doing the work. The same title can cover old duties in one workplace and new duties in another. Workers can describe similar activities with different words, and a self-reported phrase can be ambiguous. The CAIO is therefore an observable boundary around a harder question: what tasks are people actually performing, and what expertise do those tasks require?
The study also separates technology-linked new work from other new work. An artificial-intelligence specialist or a database administrator is plausibly connected to technological change. A drama therapist or a conference planner may reflect changing demand, demographics, or patterns of consumption instead. This distinction is useful because “new” does not tell us why a specialty appeared. Technology is one source of new work, not its definition.
The technology-linked classification is itself deliberately limited. The researchers used a large language model to classify titles, checked a random subsample with a human audit that showed 95% agreement, and compared the result with links to patents. They describe the measure as coarse. It can help compare broad patterns, but it cannot reveal every kind of expertise or settle whether a particular title is technologically driven in every workplace.
A vintage is therefore best understood as a dating device, not a complete biography. It tells us when a title entered an administrative index. It does not, by itself, tell us when the tasks began, how many workers performed them, whether they were difficult to learn, or whether the title will retain its wage premium.
How the study matches people to titles
The measurement becomes more informative when it uses individual occupational write-ins rather than only broad occupation codes. The historical part of the study uses the 1940 and 1950 Census Complete Count files. The contemporary part uses confidential American Community Survey data from 2011 through 2023. Both contain more detailed occupational descriptions than the three-digit occupation codes normally released in public-use files.
For each worker, the researchers compare the reported occupational write-in with candidate CAIO titles inside the same broad occupation code. In the historical records, the descriptions and titles are represented as word embeddings, and the match is based on semantic similarity. In the contemporary records, the study uses character-based fuzzy matching, which looks at shared strings or fragments. The best-scoring candidate is selected, but a new title must meet a more demanding threshold than an established title. That conservative choice reflects a sensible prior: most workers are in preexisting work, so a weak textual resemblance should not be enough to reclassify a worker as doing new work.
The logic can be summarized as a chain:
Every arrow is a possible source of uncertainty. The worker may use an unusual description. The index may not contain the right synonym. The similarity measure may favor a phrase that looks close but describes different duties. And a successful match tells us that a reported job title resembles a title in the index; it does not directly observe the worker’s full task bundle.
The historical and contemporary procedures are not identical, but the goal is the same: identify whether the worker’s reported specialty matches a title that is new relative to the relevant period. In the historical analysis, work introduced after 1930 is considered new for the 1950 data. In the contemporary analysis, new work refers to titles that emerged since 1970. The different windows are part of the result, not a detail to be discarded.
The contemporary window has another important limit. The last CAIO release available to the study corresponds to the 2018 American Community Survey. As a result, the measure does not capture titles introduced between 2019 and 2023. A claim that the study measures every new occupation present in the 2020s would go beyond its evidence. It measures new work as represented in the available title system, not all novelty that may have appeared after the classification stopped updating.
Why direct observation changes the count
Earlier research often had to estimate employment in new titles indirectly. The occupation-level method takes the share of detailed titles within a three-digit occupation that are new and multiplies it by total employment in that occupation. If ten of 200 titles are new, the method treats 5% of employment in the broad occupation as new work.
That calculation is useful when detailed worker descriptions are unavailable, but it makes a strong assumption: employment in each new title reaches its long-run proportion immediately after the title appears. In reality, a specialty may begin with a handful of workers, grow as employers discover its value, and later shrink if demand proves temporary. A title can be administratively visible before it has a large employment base.
Imputation also smooths away the geography and worker-level variation that matter to the concept of expertise. If a specialty first appears in a city or a particular industry, assigning the same title share across a broad occupation can make its opportunity look evenly distributed when it is not. The indirect method cannot then compare the education, age, earnings, or location of the people actually doing the new work with similar workers doing established work. Direct write-ins do not solve every problem, but they make those comparisons possible and allow the analysis to ask whether entry is accumulating gradually rather than appearing fully formed.
The direct write-in method avoids that immediate-steady-state assumption. It asks whether individual workers are actually reporting work that matches a new title. The study estimates that 18.3% of workers in the 2011–2023 data were in jobs introduced since 1970. In the 1950 Census data, 7.2% were in work introduced after 1930.
Those figures are not a simple historical trend. The contemporary window spans about five decades of title emergence, while the historical window spans about two. Employment in new work does not necessarily accumulate at a constant rate, and the classification systems and economic conditions differ. The estimates describe the samples and definitions used; they do not establish that exactly the same share of all workers, in every country or year, is in new work.
The contrast with imputation is revealing. The title-share method implies that roughly one worker in three is currently in new work, while the direct write-in estimate is closer to one worker in five. The imputation method overstates the overall level because it assigns the mature share of titles to employment immediately, especially for the newest cohorts. Yet it is not useless. The imputed and directly observed distributions across broad occupations are highly correlated, with a correlation of 0.91. Within individual occupations, however, the correlation of new-work shares is much lower, at 0.51. The method is better at showing where new work is concentrated than at estimating how much of each occupation is actually new.
This is a general lesson about measurement. An imperfect proxy can preserve broad structure while distorting levels. If a method correctly identifies that new work is moving toward professional and managerial occupations, that does not mean its estimate of the percentage of workers in new work is equally reliable. Distribution and prevalence are separate claims.
The historical comparison also shows why the content of new work matters. In 1950, more than four in ten workers in new titles were in production or clerical and administrative support, occupations around the middle of the earnings distribution. In the recent data, more than four in ten are in the two highest-paid broad categories, professional and managerial. Among workers with less than four years of college, recent new work has shifted toward health and personal services, while historical new work was more concentrated in production and clerical support.
These are not merely changes in labels. They suggest that the route into emerging work has changed as the economy has changed. But the method still cannot tell us everything about the tasks inside those categories. A title-based measure can reveal where a specialty is recognized and which workers report it. It cannot by itself reveal the full training process, the quality of the work, or whether the specialty is a genuinely new bundle of tasks rather than a new name for familiar activity.
What the measure can and cannot say
The strongest conclusion is not a single percentage. It is a disciplined boundary around the evidence.
The measure can identify occupational titles that entered a recognized index during a defined period. It can connect those titles to individual workers rather than assigning every worker in a broad category the same probability of doing new work. It can compare vintages, examine who reports the titles, and estimate how workers in those titles differ from workers in preexisting work.
The measure cannot prove that every title added to the index represents a wholly new task. It cannot capture post-2018 titles in the contemporary sample. It cannot remove all ambiguity from self-reported descriptions. And it cannot turn a title’s recency into a guarantee of high pay or lasting demand.
That last boundary matters when reading a job forecast. An occupation may rank near the top of a headcount table even though the relevant new specialty is hidden inside it. Conversely, a new title may be economically important without generating enough jobs to appear among the occupations with the largest numeric gains. The two sources answer different questions because they count at different levels.
The measurement strategy therefore prepares, rather than settles, the distributional question. Once new work has been identified as a classified specialty with a defined time window, we can ask who performs it, where they live, and what prior education or experience they bring. A prevalence estimate is not the final judgment about new work. It is the point at which the analysis can begin to examine access to emerging expertise.