Using newly available 1940 and 1950 Census Complete Count files and confidential American Community Survey data, we show that new work differs from simply more work in existing occupations. It draws younger and more educated workers, and it pays wage premiums that persist well beyond entry and that shrink across vintages as expertise diffuses. New work emergence traces to specific local demand shocks. New work thus counterbalances automation not only by adding employment, but by generating new domains of scarce human expertise.

The returns and risk premia of stocks with similar characteristics but different levels of ownership comove far more with shocks to the risk-bearing capacity of financial intermediaries. This implies that intermediaries are not a veil, even within the least-intermediated asset class.

We use advances in natural language processing and large language models to construct new measures of workers’ technology exposure spanning nearly two centuries. Linking these measures to Census occupation data, we show that technological progress historically increased demand for higher-educated, better-paid, and more female-dominated occupations, but our calibrated model predicts that AI will reverse these trends, favoring lower-educated, lower-paid, and more male-dominated jobs.

We examine the content of newly emerging job categories over an 80-year period and the countervailing roles of labor-augmenting and automating innovations in generating demand for new work. The distribution of new work emergence polarized from middle-paid production and clerical occupations over 1940–1980, to high-paid professional and, secondarily, low-paid services since 1980. While the demand-eroding effects of automation innovations have intensified in the last four decades, the demand-increasing effects of augmentation innovations have not.

We propose a simple and computationally tractable methodology for computing similarity between two documents along with economic applications of the method.

Using idiosyncratic stock returns, I estimate heterogeneous firm-level labor supply elasticities by labor productivity, worker skill, and time. After accounting for the mitigating impact of adjustment costs, I use these elasticity estimates to quantify how wage markdowns affect the following: a wide cross-sectional labor share spread by productivity; the public firm aggregate labor share decline from 1991-2014; and productive firms’ high profits and valuations, despite low investment. Overall, profits from wage markdowns are worth 20-25% (4%) of aggregate capital income (revenues). Productive firms’ labor market power over skilled workers plays a central role in these patterns.

We construct occupation- and firm-varing measures of workers’ task exposure from 2010 to 2023. We show theoretically that labor demand decreases in the average exposure of workers’ tasks to AI technologies, and increases in the dispersion of task exposures to AI driven by labor effort reallocating towards unaffected tasks. Additionally, firm-level productivity effects from AI use tend to raise employment overall. We document strong empirical support for all these predictions.

We develop measures of workers’ exposure to labor saving and labor augmenting technologies. Labor-saving technologies uniformly predict earnings declines for individual incumbent workers; labor-augmenting innovations only do so for skilled incumbents, but they also predict higher aggregate occupational labor demand and increased earnings for occupational entrants. We interpret our findings through a model with automation and skill displacement. Previously titled ‘Technology, Vintage-Specific Human Capital, and Labor Displacement: Evidence from Linking Patents with Occupations.’

AI models must be trained on data that workers generate, and this complementarity between labor and data changes how AI affects the labor market. Linking Census firm surveys on AI adoption to individual tax records, we find that exposed workers gain on average, especially in occupations where firms do substantial in-house AI R&D. We then estimate a directed search model in which workers generate data, firms adapt their skill weights, and workers can retrain. Even large improvements in AI fail to generate widespread displacement, with long-run unemployment bounded below 6.5%.

Using Census employer-employee matched data, we provide the first direct evidence that firms’ intangible investment raises the portable human capital of their workers, making these firms “human capital incubators.” Incubation goes together with higher profitability, more market power, and inflows of young workers, as in a model where workers’ preference for skill development gives incubators a labor market advantage. For top-tercile incubators, incubation is worth about 13% of firm value, with 57% of that value spilling over to workers and downstream employers.