Labor Market AI Exposure: What Do We Know?

This piece echoes earlier work by Frank et al. (2025), who also found disagreement between AI exposure scores. They encourage a hybridized approach instead of utilizing a single score, as “using only one AI exposure score will misrepresent AI's impact on the future of work.” This piece utilizes some of the same exposure metrics as Frank et al. but includes others in its analysis. 

Acemoglu et al. (2022) also consider several metrics in their work on labor markets, including some of the ones used in this analysis (Felten et al. and Webb). They highlight how each of the three metrics attempts to measure a different aspect of AI, and they note that their methods select for different forms of AI exposure. “The Felten et al. measure, for example, is particularly high for managers, professionals, and office and administrative staff and is very low for service, production, and construction workers, capturing the fact that these occupations involve various manual tasks that cannot currently be performed by algorithms. The Webb measure is not particularly high in sales occupations and shows a strong positive relationship with occupational wage percentiles.”

Each metric utilizes a different occupational categorization, ranging from SOC codes to altered 1990 CPS codes. To render them comparable, we crosswalk from each starting point to eventually arrive at the SOC 2018 occupational codes.

This leaves us with 867 occupations with at least one exposure score, and 710 occupations for which every metric provides a score. 

We follow Goldschlag and Eckhardt’s crosswalking method to render the disparate measures comparable. We detail each metric and its associated crosswalk below.

Felten et al. (2021) created the AIOE metric based on surveys collected from Amazon’s Mechanical Turk workers. They were asked about several AI applications’ capacity in abilities from the O*NET ability database, and Felten et al. condensed these responses into scores for the various SOC 2010 codes to which the abilities belonged. 

The crosswalk here is from SOC 2010 codes to SOC 2018 codes. For cases where multiple codes in one classification map to a single code, the abilities associated with each occupation are weighted by level of importance to the occupation as reported in the database. For cases in which an ability appears in one period and not another, its weighted importance score is filled with a zero, following the assumption that its absence indicates its unimportance. These weights are then used to calculate a similarity score between the occupations’ SOC 2010 and SOC 2018 codes, calculated by normalizing the ability scores to a unit vector and then calculating the dot product between the linked occupations. This similarity score scales the occupation’s exposure metric from one code scheme to the other, with a score of 1 indicating they are identical and a score of 0 indicating no overlap.

Eloundou et al. (2024) created the dv_rating_beta and human_rating_beta metrics, which assess whether GPTs and technologies built on top of them can speed up the time in which workers complete O*NET tasks. The human rating was generated by human annotators while the dv_rating was annotated by GPT-4. These measures are reported at the SOC 2019 level, which is a more detailed version of the SOC 2018 codes. We reduce the codes to their less detailed SOC 2018 codes and aggregate to the broader occupational category where appropriate (i.e., Chief Executives and Chief Sustainability Officers are collapsed into Chief Executives).

Eisfeldt et al. (2023) created the genaiexp_estz_total and genaiexp_estz_core metrics. The authors build on Eloundou et al.’s work, applying their own rubric assessing generative AI’s capacity to improve productivity for O*NET tasks. The “total” metric examines all O*NET tasks while the “core” metric considers only tasks labeled as such for their associated occupation. Both metrics were annotated using GPT-4. Like the AIOE scores, these exist at the SOC 2010 level and follow the same crosswalking procedure. 

Webb (2020) created the pct_ai metric, which follows a method distinct from the others. He links patent filings with O*NET occupations and tasks by finding overlap between their descriptions. He then links these matches to associated occ1990dd classification codes developed by Dorn (2009) and refined by Deming (2017). This yields a score indicating that jobs exposure to patented AI technologies. Going from the occ1990dd occupation codes requires an involved crosswalk. First, they are mapped to the 1990 Census occupation codes. Then, they are translated to 2018 Census codes, with occupations that are either collapsed or expanded being weighted by their relative presences in distinct time periods. Finally, the 2018 Census codes are mapped to the corresponding SOC 2018 codes.

Microsoft researchers Tomlinson et al. (2025) created the ai_applicability_score. Their methods closely align with both Eloundou et al. and Eisfeldt et al. in their use of O*NET tasks. They distinguish their work from other researchers by looking at individual conversations with Microsoft’s Copilot AI. They categorize conversations by the user goals, type of use, and success of that model in completing O*NET tasks, the results of which are then aggregated to the SOC 2018 level. Since this data already exists at the SOC 2018 level, there is no need to crosswalk.

Below we describe our data processing in detail.

Following crosswalking, we are left with 867 occupations covered by at least one metric, 710 covered by all metrics, and 779 covered by all but Pct AI. Given that every metric is on a different scale, they are all normalized to z-scores so that we can weight subsequent analysis using PCA weights. Z-score standardization puts every metric on the same scale so that they have a mean of zero and a standard deviation of 1. 

PCA weights refer to Principal Component Analysis, so that when calculating scores, each is weighted according to how much it contributes to overall variance. We use PCA weighted z-scores instead of ranks because it doesn’t flatten the difference between occupations. While two occupations could be next to each other using ranks, that is not a strong indication of the magnitude of the difference between them. Z-scores provide information on the magnitude of a difference instead of just ordering. 

Including Webb, the PCA weights are as follows:

Since the datasets with and without Webb drop some occupations that lack full coverage across metrics, it is worth testing whether the omitted occupations differ significantly in rank from the included ones. We perform Mann-Whitney tests both including and excluding Webb. The results indicate that there is a statistically significant difference between included and excluded occupations both with and without Webb. The excluded occupations tend to have a slightly lower exposure scores than the occupations included. As results indicate above, this is not an overly problematic finding.

Beyond the figures we show in the main body, we perform several other tests to better understand our results.

This simple test compares the ranks of DV Rating Beta against each of the other metrics. Across metrics, the trend is linear with high correlation (.7-.9). The highest correlation is with Human Rating Beta, which is unsurprising since both come from the same Eloundou et al. paper. The slope is significant for all of them, and the R-squared is between .47 and .78 for all of them, with most being at the high end of that range.

In these plots, we compare the K density of the normalized z-scores against DV Rating Beta and against the PCA weighted score of all measures.

For the plots comparing the measures against DV Rating Beta, there is a significant difference in distribution between each metric, apart from Human Rating Beta. AIOE has an even central distribution with very long tails, indicating that it has more occupations with very high and very low ratings. Both DV Rating Beta and Human Rating are peaked at the low end and moderately high end of exposure.  The GenAI metrics both have heavy right skew with long tails, showing that most occupations are centrally scored but with a meaningful and widespread amount of highly exposed occupations. AI Applicability has a high single peak centered around -1.

The regressions included in the main text are multivariate regressions of the variance of the measures against the PCA weighted score across those same metrics while controlling for share of occupation’s workers that are women, the log of salary, and the occupation’s major occupation category. This tests whether the variance (understood here at the disagreement on scores between metrics) varies with the scores themselves. We also test both model specifications.

This lets us answer two different questions. In specification 1, we test whether the metrics agree across all levels of exposure. In specification 2, we test whether DV Rating Beta is a good predictor of the disagreement in other metrics.

The key independent variables are highly significant for both model specifications. As with the simple univariate test, variance increases as exposure gets higher. This is a statistically significant change, with the coefficient indicating that a 1 unit increase in PCA score leads to a 0.0617 increase in variance, which suggests the increase in disagreement is real but only moderate. While the effects of the controls are statistically insignificant, the results for the main independent variables are robust to the inclusion of the controls.