GPTs are GPTs: Labor market impact potential of LLMs

Eloundou et al. (2024) estimate that around 19% of the total U.S. workforce could see at least half of their tasks affected by large language models alone.

“Eloundou, Manning, Mishkin, and Rock (2024) estimate that around 19% of the total U.S. workforce could see at least half of their tasks affected by large language models (LLMs) alone.”

Eloundou et al. (2024) construct an exposure measure by rating whether access to such technologies could reduce the time required to complete a given task by at least 50 percent.

“Targeting LLM capabilities specifically, Eloundou et al. (2024) construct an exposure measure by rating whether access to such technologies could reduce the time required to complete a given task by at least 50 percent.”

Under the Eloundou index, the 22-25 ChatGPT coefficient is again negative, smaller in magnitude (-0.066), and not statistically significant.

“The 22-25 Chat-GPT coefficient is again negative but smaller in magnitude (-0.066) and not statistically significant.”

The 31-34 headcount and wage cells that were significant at the 5% level under Massenkoff become +0.020 (p = 0.29) and +0.012 (p = 0.21) under Eloundou, and the 50+ wage cell becomes +0.008 (p = 0.24).

“the 31-34 headcount and wage cells that were significant at the 5% level under Massenkoff and McCrory (2026) (Table 2) become +0.020(p = 0.29) and +0.012 (p = 0.21) respectively, and the 50+ wage cell becomes +0.008 (p = 0.24).”

The 22-25 headcount cell under Eloundou is again negative (-0.024) but not statistically distinguishable from zero.

“The 22-25 headcount cell is again negative (-0.024) but again not statistically distinguishable from zero.”

None of the fourteen real-treatment cells reaches conventional significance under the Eloundou index.

“None of the fourteen real-treatment cells reaches conventional significance under Eloundou”

The four occupational AI exposure indices are measured on different scales, ranging from a standardized score with mean near zero and standard deviation near one to a right-skewed usage-weighted share.

“The four indices live on different scales -Felten et al. (2021) delivers a standardized score with mean roughly zero and standard deviation roughly one; Pizzinelli et al. (2023) delivers a positive, bounded score on a wider range;”

Under both the Massenkoff and McCrory (2026) and Eloundou et al. (2024) indices, real-treatment ATTs on headcount are small and positive in the all-ages and below-40 bands, while the 2019H2 backdating placebos are systematically larger and statistically significant in nearly every cell.

“The pattern is consistent across both indices: real-treatment ATTs are small and positive on headcount in the all-ages and below-40 bands, while the 2019H2 backdating placebos are systematically larger in magnitude and statistically significant in nearly every cell.”

The 22-25 panel under Eloundou mirrors the Massenkoff result: pre-treatment coefficients are flat (consistent with the clean joint pre-trend test) and post-treatment coefficients drift modestly negative but are not statistically distinguishable from zero.

“The 22-25 panel mirrors the Massenkoff result: pre-treatment coefficients are flat (consistent with the clean joint pre-trend test) and post-treatment coefficients drift modestly negative but are not statistically distinguishable from zero.”

The 50+ panel under Eloundou reproduces the persistent positive trajectory seen under Massenkoff on a similarly dirty pre-trend.

“The 50+ panel reproduces the persistent positive trajectory seen under Massenkoff on a similarly dirty pre-trend”

Theoretical AI exposure is measured using the Eloundou et al. (2024) classification.

“For theoretical exposure, we use the measure developed by Eloundou et al. (2024)”

The estimated interaction between summer 2026 and the theoretical AI exposure measure of Eloundou et al. is small and not statistically significant in either the standard or the expanded specification.

“| Summer 2026 × AI exposure | 0.009 (0.008) | 0.012 (0.009) | 0.008 (0.008) | 0.012 (0.009) |”

The rubric was applied to each DWA and job task by a team of human labelers, and results were aggregated at the occupational level to produce an overall score representing the proportion of job tasks within an occupation that can achieve significant time savings by utilizing an LLM.

“The rubric was applied to each DWA and job task by a team of human labelers. Results were subsequently aggregated at the occupational level to produce an overall score representing the proportion of job tasks within an occupation that can achieve significant time savings by utilizing an LLM.”

Higher wages are associated with greater occupational exposure to LLM.

“Our univariate findings align with past studies showing a relationship between higher wages and greater occupational exposure to LLM (18, 22).”

LLM can be used to summarize and synthesize vast amounts of information, manage customer inquiries, support organizational decision-making and generate technical and creative content.

“LLM can be used to summarize and synthesize vast amounts of information (eg, legislative analysis tools), manage customer inquiries (eg, chatbot customer service applications), support organizational decision-making (eg, resume screening and job matching platforms) and generate technical and creative content (eg, content ideation tools) (17–21).”

There is a small body of research which signal an inverse relationship between occupations that are precarious as well as exposure to LLM.

“There is a small body of research which signal an inverse relationship between occupations that are precarious as well as exposure to LLM (ie, where LLM can be used to perform tasks that make up an occupation).”

Eloundou et al created a rubric to assess occupational LLM exposure based on the share of an occupation's job tasks where access to an LLM alone or with a simple interface would lead to 50%-time savings and the share of an occupation's tasks where additional software is needed on top of an LLM to realize the 50%-time savings.

“Using this definition, they created a rubric to assess occupational LLM exposure based on the share of an occupation’s job tasks where access to an LLM alone or with a simple interface would lead to 50%-time savings (based specifically on manual human coding) and the share of an occupation’s tasks where additional software is needed on top of an LLM to realize the 50%-time savings.”

We used the novel LLM exposure measure by Eloundou et al to examine the impact of the technology on occupations.

“We used the novel LLM exposure measure by Eloundou et al (18) to examine the impact of the technology on occupations.”

Occupational exposure to LLM was highest among occupations that were characterized by higher wages.

“Occupational exposure to LLM was highest among occupations that were characterized by higher wages (one indicator of whether an occupation is precarious or not) (18).”

Eloundou and colleagues recently found that just under 20% of occupations in the US consisted of job tasks in which 50% are exposed to LLM.

“Eloundou and colleagues (18) recently found that just under 20% of occupations in the US consisted of job tasks in which 50% are exposed to LLM.”

Over time, the uneven exposure of occupations to LLM and other emerging AI systems may contribute to labor market inequities and health disparities.

“Over time, the uneven exposure of occupations have to LLM and other emerging AI systems may contribute to labor market inequities and health disparities (18, 42).”

In the absence of direct measures, the study utilized a novel proxy measure of the extent to which an occupation is composed of job tasks that can be performed by an LLM.

“In the absence of direct measures, we utilized a novel measure which offered a proxy regarding the extent to which an occupation is composed of job tasks that can be performed by an LLM (18).”

AI agents are increasingly deployed to support everyday workplace tasks.

“They are increasingly deployed to support everyday workplace tasks (Eloundou et al. 2024;Xi et al. 2025;Mohney 2025).”