They aren't just looking at tech jobs in the study but all job in the data set that they've coded as being AI exposed.
> We use two different approaches for measuring occupational exposure to AI. The first uses exposure
measures from Eloundou et al. (2024). Eloundou et al. (2024) estimate AI exposure by ONET
task using ChatGPT validated with human labeling. They then construct occupational exposure
measures by aggregating the task data to the 2018 SOC code level. We focus on the GPT-4 based
β exposure measures from their paper.
>The second primary approach we take uses data on generative AI usage from the Anthropic
Economic Index (Handa et al., 2025). This index reports the estimated share of queries pertaining
to each ONET task based on a sample of several million conversations with Claude, Anthropic’s
generative AI model. It then aggregates the data to the occupational level based on these task
shares. One feature of the Anthropic Economic Index is that for each task it also reports estimates
of the share of queries pertaining to that task that are “automative,” “augmentative,” or none of
the above. We use this information as an estimate of whether usage of AI for an occupation is
primarily complementary or substitutable with labor.14
> We use two different approaches for measuring occupational exposure to AI. The first uses exposure measures from Eloundou et al. (2024). Eloundou et al. (2024) estimate AI exposure by ONET task using ChatGPT validated with human labeling. They then construct occupational exposure measures by aggregating the task data to the 2018 SOC code level. We focus on the GPT-4 based β exposure measures from their paper.
>The second primary approach we take uses data on generative AI usage from the Anthropic Economic Index (Handa et al., 2025). This index reports the estimated share of queries pertaining to each ONET task based on a sample of several million conversations with Claude, Anthropic’s generative AI model. It then aggregates the data to the occupational level based on these task shares. One feature of the Anthropic Economic Index is that for each task it also reports estimates of the share of queries pertaining to that task that are “automative,” “augmentative,” or none of the above. We use this information as an estimate of whether usage of AI for an occupation is primarily complementary or substitutable with labor.14