What We Do and Don't Know About How AI is Affecting the Labor Market

If we want to understand AI impacts, why isn’t it sufficient to simply compare exposed and unexposed workers’ outcomes today? The reason is that these occupations (and their workers) are different in ways that are unrelated, or only incidentally related, to AI exposure. 

According to the standard metrics, which Budget Lab has analyzed in detail previously, AI exposure is distributed unequally across workers.1 New large-language models are very capable of assisting with coding activities, for example, but generally cannot intervene in the physical world and have limited utility for tasks requiring such intervention. Distinguishing different types of jobs along this dimension is in principle useful.

However, our AI exposure metric and others currently in use have some important limitations. These measures tend not to be specific about whether AI is likely to replace or enhance a worker’s labor, or to what extent the technology is currently being used (Massenkoff and McCrory 2026). AI exposure metrics are ultimately informed guesses as to which tasks and occupations are most likely to be affected by AI tools, based on some combination of human and LLM judgment as to which tasks overlap with LLM capabilities. This can change over time and could be wrong from the start.

Those caveats in mind, it is remarkable how unequally distributed AI exposure is across the labor market. Workers in exposed jobs have persistently different outcomes than their counterparts, as shown in Table 1. We consider occupations to be exposed to AI when they are in the top third of occupations by the Budget Lab’s summary metric, and unexposed when they are in the bottom third.2 As measured by number of distinct occupations, the most common major occupational group in the exposed category is Office and Administrative Support; in the unexposed category, the most common group is Protective Services.

Workers in exposed occupations are more likely to be women and much more likely to have a four-year college degree. Based on pre-pandemic data, they are also less exposed to the business cycle, suffering smaller reductions in employment during recessions. Larger coefficients indicate a more procyclical pattern, with employment rising during booms and falling during downturns. (See the appendix for more discussion of this.) Exposure is also correlated with other factors which may affect hiring and firing, like remote work (Kolko 2026).

These gaps make it inherently difficult to interpret comparisons between exposed and unexposed occupations.3 Fortunately, there exist econometric strategies that can be useful for handling at least some of these differences, whether observed or unobserved in our data. Our preferred strategy, synthetic differences-in-differences, is an amalgam of the better-known synthetic control and differences-in-differences methods. It allows one to construct an apples-to-apples comparison group for AI-exposed workers using a weighted combination of unexposed occupations. Differences between this comparison group and exposed workers can, under certain assumptions, be interpreted as effects of AI.4 Most importantly, one must assume (as discussed above) that AI exposure metrics have some ability to distinguish, however imperfectly, occupations that are more or less likely to be affected by AI.

We now apply this strategy by showing the average employment share of an occupation (out of the civilian population) in the AI-exposed group against the average share of the synthetic comparison group. Figure 1 shows these shares over time, with the dashed vertical line indicating the fourth quarter of 2022, when ChatGPT was released for a mass audience.5

Because SDID has the desirable feature of accommodating level differences in the AI-exposed and unexposed groups, it can be difficult to visually determine the difference in the two groups: the estimated impact of exposure to AI. Figure 2 shows that difference, along with confidence intervals that indicate whether the impact in a given quarter is significantly different from zero.6 No impact is evident through the post-2022 time window.7

We also examine the log real hourly wages of AI-exposed and unexposed workers.8 Figure 3 shows these trends, and Figure 4 shows the impact of LLM introduction. Here again, we see no statistically or economically significant impact. 

In appendix figures, we additionally show results for the unemployment rate, for our entire sample and for a subgroup of workers ages 16 to 34. These figures indicate positive effects in the most recent quarter—roughly half a percentage point increase in the entire sample, and more for the 16-34 year old subsample—but both are statistically insignificant as of the first quarter of 2026. It is possible that AI has raised unemployment for workers in exposed occupations; however, unemployment may not be the ideal outcome to track for this purpose. Unemployment effect estimates are especially imprecise in our context and subject to the limitation that occupation measurement is complicated for unemployed workers, many of whom are not associated with a specific occupation. 

A small dose of economic theory is helpful for interpreting the results above. In the labor market, there are four basic patterns that could characterize employment and wages, each of which is associated with a different kind of labor market “shock”. The possibilities are:

1. Employment rises and wages fall: a positive supply shock
2. Employment falls and wages rise: a negative supply shock
3. Employment and wages both rise: a positive demand shock
4. Employment and wages both fall: a negative demand shock

The pattern that many anticipate, for AI-exposed workers, is the last one: a negative demand shock. As AI tools allow employers to automate certain tasks and the workers who perform them, one could expect employer demand for labor to fall, and both employment and wages to decline. 

Within the synthetic differences-in-differences framework, we do not yet see evidence of this pattern. However, there are some important caveats that should be kept in mind. First, LLMs are constantly improving, such that the economic impacts from early models seem likely to be smaller than those of recent models. Second, the validity of the SDID approach taken in this analysis depends on whether the introduction of LLMs can be distinguished from other events, like the rise in remote work, that are closely correlated with exposure to AI.9 It also depends on how well AI-exposure metrics proxy for the tasks and jobs actually at risk in the age of AI. If “AI-exposed” occupations include some that receive negative demand shocks and some that receive positive shocks, our analysis would miss the former cause as it averages across the whole group.

Finally, it worth noting that the Bureau of Labor Statistics’ Current Population Survey, our data source throughout this analysis, is best suited to analysis of broad groups and somewhat underpowered for subgroup analysis, like the 22–27 year old recent college graduates that have been a special focus in some AI labor market impact discussions.10 If labor market effects of AI are currently limited to a narrow slice of the workforce, other datasets and research designs would be better suited to identify them. 

Using this and other econometric approaches, and keeping in mind their diverse strengths and weaknesses, the Budget Lab will continue to monitor the labor market for signs of AI effects.