Automation Experiments and Inequality
Seth Benzell Kyle R. Myers Chapman, MIT, & Stanford Harvard & NBER · 2025
Read Automation Experiments and Inequality by Seth Benzell Kyle R. Myers Chapman, MIT, & Stanford Harvard & NBER (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-24923.txt.
Introduction
Seth Benzell and Kyle Myers frame what lab tests of new work tech can and can not say on gaps between workers. Many such tests ask if a tool lifts high skill or low skill staff more, then use that to guess long run gaps. Their model has each worker do a task or hand it to the tech.
The gap effect turns on two things: how alike task skills are across staff, and how staff skill stacks up next to the tech. As the tech gets strong, the sign of the gap effect can flip more than once. So a test at one tech level may point the wrong way for the next. Their frame warns firms to read such tests with care.
Catalog record
| CREATOR | Seth Benzell Kyle R. Myers Chapman, MIT, & Stanford Harvard & NBER |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-24923
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Automation Experiments and Inequality Seth Benzell Kyle R. Myers Chapman, MIT, & Stanford Harvard & NBER Abstract An increasingly large number of experiments study the labor productivity effects of automation technologies such as generative algorithms. A popular question in these experiments relates to inequality: does the technology increase output more for high- or low-skill workers? The answer is often used to anticipate the distributional effects of the technology as it continues to improve. In this paper, we formalize the theoretical content of this empirical test, focusing on automation experiments as commonly designed. Worker-level output depends on a task-level production function, and workers are heterogeneous in their task-level skills. Workers perform a task themselves, or they delegate it to the automation technology. The inequality effect of improved automation depends on the interaction of two factors: ( i ) the correlation in task-level skills across workers, and ( ii ) workers’ skills relative to the technology’s capability. Importantly, the sign of the inequality effect is often non-monotonic — as technologies improve, inequality may decrease then increase, or vice versa. Finally, we use data and theory to highlight cases when skills are likely to be positively or negatively correlated. The model generally suggests that the diversity of automation technologies will play an important role in the evolution of inequality. Introduction Recent advances in the capabilities of algorithms have led to a surge of investigations into the economics of automation technologies. An important, open question is the effect of these technologies on earnings inequality. Researchers have approached this question using both macroeconomic and microeconomic methods. In this paper, we are interested in what can and cannot be learned from the microeconomics. 1 1 1 The macroeconomic work typically asks: after general equilibrium effects play out, will the new distribution of jobs and earnings be less or more unequal? The answer appears often is that technologies that reduce demand for certain types of labor in the short run may increase demand for that labor in the long run, and vice-versa. For overviews covering historical and current waves of automation, see: Katz and Autor 1999 ; Acemoglu and Autor 2011 ; Acemoglu and Restrepo 2019 . Specifically, we focus on the microeconomics of automation experiments: studies that use exogenous variation in the availability of a technology to estimate productivity effects. These experiments are valuable because they yield internally valid estimates (typically, in partial equilibrium) and provide managers and policymakers information about technologies at a relatively high frequency. In order to estimate inequality effects, experimenters often test whether the technology increases the output of higher- or lower-skilled workers with the same job. 2 2 2 For examples of automation trials that report inequality results, see: Noy and Zhang 2023 ; Chen and Chan 2024 ; Choi, Monahan, and Schwarcz 2024 ; Kreitmeir and Raschky 2024 ; Cui et al. 2025 ; Brynjolfsson, Li, and Raymond 2025 ; Dell’Acqua et al. 2025 ; Kanazawa et al. 2025 ; Kim et al. 2025 ; Otis et al. 2025 ; Roldán-Monés 2025 . In June of 2025, The Economist summarized the inequality effects from the automation experiments to date: “ Although early studies suggested that lower performers could benefit … newer studies look at more complex tasks… In these contexts, high performers benefit far more than their lower-performing peers, ” which implies that the skill level of the workforce is an important feature and suggests a future of skill-biased technical change, as implied by the article’s title “ How AI will Divide the Best from the Rest, ” ( The Economist 2025 ). But what exactly do we learn if we see an automation experiment lead to more, or less, inequality? What features of the workforce are relevant to the inequality effect? If a technology increases inequality now, will it continue to increase inequality as it becomes more capable? To answer these sort of questions, we provide a new model of automation experiments where the occupations and workforce are held fixed. The model is in the task-based tradition; output is increasing in task-level inputs. However, unlike many macroeconomic models of task-based production, we explicitly model jobs as production functions at the individual level . Individual workers are defined by their endowments of skill in each task, which motivates our focus on a feature of the workforce that is rarely highlighted: the correlation of workers’ skills across tasks. This correlation describes, for example, whether the most analytical physicians are also the most empathetic, or whether the best fielder on the baseball team is also the best batter. We study automation experiments involving either positive or negatively correlated skills among workers. Automation technologies can perform one (or more) of the tasks in production, and a technology’s capability may be below or above a given workers’ skill. We assume rational expectations, so workers only automate (i.e., delegate to the technology) if the technology’s capability is above their skill. 3 3 3 This assumption is not innocuous, as some automation experiments do find workers to perform worse than placebo when given access to generative algorithms ( Dell’Acqua et al. 2025 ; Otis et al. 2025 ), which suggests the presence of biased beliefs. The model provides clear predictions about the change in inequality (i.e., the difference in output between worker types) given a marginal increase in an automation technology’s capability. Even in simple settings with only two tasks, perfect skill correlations, and single-task automaton, we obtain results that may seem counter-intuitive at first, but become very clear through the lens of the model. Changes in inequality do not depend on the absolute level of workers skill, as the The Economist quote above implies or as intuition might suggest. It is the interaction of skill correlation and technological capability that determines the inequality effect. For example, consider two samples of workers with identical, two-task Cobb-Douglas production functions and identical, task-levelskill distributions. But in one sample, skills are positively correlated and, in the other, skills are negatively correlated across workers. Here, an increase in the capability of a technology that automates one task can decrease inequality in one sample while increasing inequality in the other. To illustrate the intuition of this result, consider Figure 1 , which reflects this example. In both cases, the production function and the initial level of inequality (per the difference in output between high-type H and low-type L ) is the same; however, in Panel A, skills are positively correlated while in Panel B skills are negatively correlated. As, for instance, Task 1 automation technology improves, the first user will be the low type in the positive correlated case, but it will be the high type in the negative correlation case. Thus, at least initially, the low type will benefit first and inequality will decrease in the Panel A case, but the high type will benefit first and inequality will increase in the Panel B case. figure 1 : Initial Changes in Output after Automation A Positive skill correlation Lo Hi B Negative skill correlation Lo Hi Note: Plots isoquants of production possibilities, highlighting the isoquant for a worker with higher output ( Hi ) and a worker with lower output ( Lo ) in two scenarios, where task-level skills are