Estimating AI exposure is useful for research but cannot yet tell policymakers the whole story of AI’s impact on the labor market

Key takeaways

• While estimates of AI exposure—calculations of how artificial intelligence may interact with certain jobs or tasks—are valuable in the absence of reliable public data sources, there are limitations to using these numbers, particularly to inform policymaking. Most studies that link AI exposure to employment outcomes, for example, cannot isolate its impact from other factors influencing the economy.
• The validity of AI exposure scores depends on the quality of the methods used to calculate them. They also carry the assumptions made during the estimation process, such as how jobs are broken into tasks or whether a chatbot can accurately assess its own capabilities.
• What this means for growth: A well-functioning U.S. labor market is essential for U.S. economic growth. In the face of potential employment disruptions from increasing use of AI across sectors of the economy, policymakers need reliable evidence to inform and properly target solutions. It is therefore vital that they understand what AI exposure scores can tell them and where the scores fall short.

Overview

Artificial intelligence has the potential to substantially change the nature and availability of work. To effectively manage the transitions this technology might bring to the U.S. economy, policymakers need information on how AI interacts with firms and workers.

Currently, there are no detailed measurements of how workers are using AI and the role that firms play in its implementation in the workplace. In the absence of these quantifications, researchers have developed a useful but imperfect proxy: the AI exposure score.

AI exposure scores offer an approximation of AI’s impact on workers and workplaces. But stakeholders who wish to use these estimates to inform policy decisions must be careful. AI exposure scores are not measurements; they are conjecture. They cannot be used to isolate AI’s role in the labor market and are not immune to bias in how they are calculated.

This column looks at how AI exposure scores can be useful for researchers before describing the limitations of these estimates and why policymakers should proceed cautiously.

AI exposure scores fill an ongoing gap in data

AI exposure can broadly be thought of as the potential for AI to interact with a job. Researchers have sought to estimate AI exposure through a variety of sources, including expert estimates, the self-assessments of AI models, resume and job description data, and the rates at which chatbots are used. The most popular calculation methods are fully explained in Equitable Growth’s synthesis of current AI and labor market research and its companion living database.

While such measures are valuable approximations, particularly for research purposes, they are limited by the assumptions and methodologies used to calculate them. As these measures are increasingly adopted by policymakers, businesses, and the media to inform and make decisions, any findings on AI and the U.S. labor market that these exposure scores imply must be interpreted with care. 

AI exposure scores are useful for researchers but have limitations

A scan of headline studies shows that AI-exposure-based research describes positive, negative, and neutral impacts of AI on employment outcomes. This variation is unsurprising, given the differences in how exposure scores are calculated. Some studies, for example, seek firm-based outcomes, while others look at worker-level effects. It therefore stands to reason that each study will yield results that are specific to its own framework.

What these studies have in common is an acknowledgement of a primary limitation: It is challenging to isolate the impact of AI from other factors influencing the economy. One widely discussed paper, for example, finds that AI exposure is related to stagnant or negative employment growth for early career workers. But its authors warn that they cannot definitively say that AI caused this negative employment growth. Indeed, in the latest version of their paper from February 2026, the authors acknowledge that employment declines that occurred before 2024 are likely associated with macroeconomic factors rather than AI.

The California Unemployment AI Tracker, or CAIT, a tool created in partnership with the California Employment Development Department and the California Policy Lab, presents a similar case. CAIT finds that Unemployment Insurance claims increased for workers highly exposed to AI who were highly educated or worked in San Francisco or in the technology sector. But its creators caveat those findings, saying that the patterns shown in the CAIT figures could also reflect the impacts of other influences on the economy, not just AI.

Likewise, an analysis published by the Yale Budget Lab does not find evidence that large language models such as ChatGPT are affecting either U.S. employment or wages. The analysis also underscores that the strength of the study’s findings depends on whether the introduction of these tools can be distinguished from other labor market shocks.  

The quality of AI exposure scores—and the research they inform—depends on the methods used to calculate them

Researchers are careful not to link AI exposure directly to any employment outcomes they find. With this caveat, their work offers some insight into the developing and consequential relationship between AI and the labor market.

Yet readers of these studies should not only be aware of the empirical limitations of each study, but also consider the assumptions underlying them. Researchers who use AI exposure estimates do so under the premise that the particular score or scores accurately reflect some dimension of AI’s interaction with work. This assumption is not always true, however, and because these predictions cannot yet be compared to evidence, it is impossible to know whether current AI exposure estimates are correct.

A few established AI exposure scores have been calculated via a method that has caused some concern among researchers. A working paper released by Michelle Yin and Hoa Vu at Northwestern University and Claudia Persico at American University demonstrates that AI exposure scores estimated with the help of large language models fluctuate substantially depending on the model used in the evaluation process. The authors warn that any bias in an AI model bleeds into the exposure scores it produces, which can distort the results of any analysis that uses the estimates.

This finding emphasizes how vital it is to be keenly aware of how, by whom (or what), and under which assumptions AI exposure scores have been calculated. These estimations (and the choices made during their calculations) are necessary to predict how AI is affecting or could affect the labor market—but they are not forgone conclusions.

How can policymakers make informed policy decisions about AI?

To ensure that the opportunities AI offers the economy are not offset by the dangers it may pose to the workforce, policymakers need to fully and accurately understand its impact on the U.S. labor market. For this, they need alternative AI impact measurements that represent more than just estimates. Such work is necessary for policymakers to understand how AI is actually reshaping the labor market so they can respond accordingly.

In the current absence of better data, AI exposure scores can reveal some useful approximations of where AI is entering and influencing the labor market and can inform potential targeting of government support. The work of those who develop AI exposure estimations in the absence of reliable public sources such as government data is indispensable. But the limitations of these numbers—and the conclusions they inform—should also be made very clear.

---

Did you find this content informative and engaging?
Get updates and stay in tune with U.S. economic inequality and growth!

Stay updated on our latest research