This website uses cookies.

By clicking the "Accept" button or continuing to browse our site, you agree to first-party and session-only cookies being stored on your device to enhance site navigation and analyze site performance and traffic. For more information on our use of cookies, please see our Privacy Policy.

  1. Home
  2. ASSA Annual Meeting
  3. 2026
  4. Program
  5. Artificial Intelligence
« Back to Results

Artificial Intelligence

Lightning Round Session

Sunday, Jan. 4, 2026 10:15 AM - 12:15 PM (EST)

Philadelphia Convention Center, 204-C
Hosted By: American Economic Association
  • Chair: Sayorn Chin, Lafayette College

The AI Shift: Gender Roles and the Changing Labour Market

Arjunan Subramanian
,
University of Glasgow

Abstract

AI technologies are transforming production and employment across sectors, yet their impact on agriculture remains understudied. Historically, mechanization has driven agricultural productivity gains while displacing workers (Olmstead & Rhode, 2001). AI, described as a transformative technology (Brynjolfsson & McAfee, 2017), is poised to reshape agricultural labour markets, particularly in developing countries with low productivity.

This paper examines AI-powered agricultural extension services through a randomized controlled trial (RCT) in India, assessing their impact on labour demand and farm productivity. Farming households, engaged in both agricultural and nonfarm work, allocate labour based on gendered roles. AI-driven advisory tools influence these allocations, altering labour demand across cultivation stages.

Three key findings emerge. First, AI-powered interventions increased paddy yields by 10%, raising profits per acre by 31%. Household income from nonfarm work declined by 65%—83% for men, 139% for women—but total income rose by 32%.

Second, AI-driven services significantly increased agricultural labour demand, particularly for women. Female labour use rose by 104% (82 additional workdays), with a 126% increase in family female labour and an 86% rise in hired female workers. However, despite greater labour input, productivity per worker declined, suggesting that AI enhances agricultural output but not individual labour efficiency.

Third, the total days worked remained unchanged, as reductions in nonfarm work were fully reallocated to farming. This shift lowered production costs, redefined gender roles in agriculture, and improved overall crop productivity.

This study contributes to three research areas: (1) AI’s economic impact on agricultural labour markets (Acemoglu et al., 2022); (2) AI adoption in agriculture and its effect on labour allocation (Fabregas et al., 2019); and (3) gendered labour outcomes in agricultural transformation (Doss, 2018). By addressing gaps in AI’s role in rural labour markets, this research offers new insights into its influence on farm productivity and employment transitions.

Memory and Generative AI

Xingjian Zheng
,
Shanghai Advanced Institute of Finance

Abstract

Generative AI is increasingly being used as economic agents. However, we know very little about their financial decision-making rules. Exploiting a novel experimental setting, we show that it uses memories to make decisions, regardless of whether the memories align with the current decision domain. When cued with images with positive emotional content, it makes riskier choices, even if it can form perfectly Bayesian beliefs. This mechanism is further causally supported with a supervised fine-tuning technique known as knowledge injection that can edit the language model's memories. Empirical analysis shows that this memory-driven behavior substantially impacts the AI agent's investment decisions and return predictability, creating significant upward or downward biases that correspond to the valence of its memories. Finally, we develop a memory-based model to explain the investment behavior of GAI agents.

What Can Artificial Intelligence Tell Us About Racial and Ethnic Groups in the U.S.?

Shiyi Chen
,
SUNY-Oneonta
Kenneth A Couch
,
University of Connecticut

Abstract

This research makes use of Current Population Survey (CPS) data to examine whether Artificial Intelligence (AI) methods group the diverse set of individual reports of racial and ethnic affiliation into those used by the federal government and most researchers. While individuals are allowed to report multiple racial and ethnic affiliations in the CPS, it is standard practice to group people based on these categories into recommended and commonly accepted categories. Here we first make use of a descriptive AI method, Multiple Correspondence Analysis (MCA), to examine which combinations of underlying reports of individuals affiliations best explain the observed variation in survey responses. Secondly, we make use of another AI method, K-Means, to group individuals based on their survey responses into mutually exclusive categories again comparing them to common practice. Finally, we perform an Oaxaca decomposition of wage outcomes based on standard race/ethnicity groupings versus those taken from the K-Means procedure to understand the implications for research on inequality.

Artificial Intelligence and the Brain: Is Innovation Getting Easier?

Buyuan Yang Sr.
,
Central University of Finance and Economics
Danxia Xie
,
Tsinghua University

Abstract

We develop an endogenous growth model in which artificial intelligence (AI) and the brain generate innovation. AI refines existing knowledge into a usable base, and the brain then recombines it into new ideas. AI's synthesis of information alleviates the brain's knowledge burdens but may weaken knowledge spillovers. Consequently, knowledge creation responds nonmonotonically to AI efficiency: at a modest level of AI efficiency, innovation slows down despite faster AI growth. Faster AI progress raises long-run growth, but its effects on research productivity and the R&D labor share are ambiguous. Finally, we characterize the conditions under which AI makes innovation easier.

The Transformative Role of Artificial Intelligence and Big Data in Banking

Junjie Xia
,
Central University of Finance and Economics and Peking University

Abstract

Leveraging a comprehensive dataset of over 4.5 million loans from a leading Chinese commercial bank and using a policy mandate for advanced financial technologies adoption as an exogenous shock, we find that the adoption of AI and big data significantly improves credit rating accuracy and loan performance. While the initial adoption of AI alone yielded modest improvements, the second-stage integration of big data analytics accounted for the bulk of the improvements, suggesting that data richness unlocked AI’s full potential. We also identify significant heterogeneity: improvements are especially pronounced for uncollateralized and short-term loans, borrowers with incomplete financial records, first-time borrowers, long-distance borrowers, and firms located in economically underdeveloped or linguistically diverse regions. These findings underscore the synergy between big data and AI, demonstrating their joint capability to alleviate information frictions and enhance credit allocation efficiency

AI Unboxed and Jobs: A Novel Measure and Firm-Level Evidence from Three Countries

Erik Engberg
,
Örebro University and Ratio
Holger Görg
,
Kiel Institute for the World Economy
Magnus Lodefalk
,
Örebro University
Natália Monteiro
,
University of Minho
Giuseppe Pulito
,
Rockwool Foundation Berlin

Abstract

This study examines the impact of artificial intelligence (AI) advancements on firm-level labor demand, utilizing granular administrative data from Denmark, Portugal, and Sweden over two decades. We introduce dynamic, time-variant indices measuring occupational exposure to AI across subdomains such as language modeling and image comprehension. These indices combine detailed AI capability data with occupational descriptions to assess workers’ susceptibility to AI advancements.

Our analysis reveals that AI exposure is primarily concentrated in white-collar occupations, particularly those requiring high cognitive complexity, analytical reasoning, and specialized knowledge, especially when social interaction is limited. However, exposure varies across AI subdomains; for instance, blue-collar workers are notably susceptible to technologies enabling remote-controlled operations.

Applying our AI exposure indices to linked employer-employee data for all non-financial firms and employees in the three countries—anchored in pre-sample workforce compositions—we find that increased AI exposure is positively associated with high-skilled white-collar employment and negatively associated with low-skilled white-collar employment. Exposure to language and image-related technologies correlates positively with demand for high-skilled white-collar work and negatively with blue-collar employment, while other AI areas exhibit varied effects across worker groups. These patterns are broadly consistent across the three countries, though associations are less pronounced in Sweden, potentially reflecting structural differences in labor markets and industry compositions.

We also observe an up-skilling effect, with the proportion of highly skilled workers increasing alongside AI exposure. This trend is most pronounced in Portugal, possibly due to differences in industrial composition or baseline workforce characteristics.

In sum, this study contributes to the literature on AI’s labor market effects by providing nuanced, dynamic measures of AI exposure across subdomains and applying them to firm-level data across multiple countries. Our findings underscore the importance of recognizing AI as a diverse set of technologies with varying employment implications depending on tasks, skills, and labor market characteristics.

Using Natural Language Processing Approach to Measure Nationalism Sentiment and Its Consequences

Kayhan Koleyni
,
Wagner College

Abstract

While nationalism is on the rise, cross-country research measuring nationalism on the individual level is very limited in the literature, and the existing measures suffer from theoretical and methodological drawbacks. In this paper, we develop a novel method to measure nationalism sentiment among 70 selected countries by using machine learning and natural language processing techniques. Then we use this new measure to investigate the impact of recent wave of nationalism on countries openness and consequently to calculate the welfare losses resulted from reduction in openness. Our results support moderate reduction in openness but sharp welfare losses among the countries with high nationalism sentiment.

AI and the Dynamics of Conflict: Theory and Experimental Evidence

Youngseok Park
,
Konkuk University
Kyu-Min Lee
,
Korea Advanced Institute of Science and Technology
Jean Paul Rabanal
,
University of Stavanger
Euncheol Shin
,
Korea Advanced Institute of Science and Technology
Sora Youn
,
Korea Information Society Development Institute

Abstract

We study how large language model (LLM) advice shapes human behavior in strategic conflict. In a laboratory experiment based on Baliga and Sjöström’s (2012) conflict game, participants choose cutoff strategies in repeated conflict games with and without AI-generated recommendations. The AI’s advice is derived from human play in the games without AI, serving as a common signal under strategic uncertainty. Comparing human-only and AI-assisted decisions, we test whether AI guidance promotes coordination and efficiency or induces new forms of distortion. The findings illuminate how human-AI interaction alters equilibrium behavior in environments of conflict and cooperation.

Algorithmic Policing

Ranae Jabri
,
University of Sydney

Abstract

What are policy-relevant tradeoffs involved in directing police presence to targeted places? This paper examines the local effects of increases in police presence induced by predictive policing, and in particular, how it affects racial disparities in outcomes. I collect a highly local novel dataset on predictive policing targeted areas, crime, and arrest from a major US jurisdiction using predictive policing. Using a natural experiment research design, I estimate that algorithm-induced police presence decreases serious violent and property crimes, but exacerbates racial disparities in arrests in traffic incidents and serious violent crimes. The evidence suggests a threefold increase in arrests of Black motorists when a neighborhood is targeted in comparison to when it is not. Local police presence can prevent crime at the cost of increasing racial disparities in arrests.

CEO Trustworthiness and Corporate Innovation: The Face Value of CEOs

Po-Hsuan Hsu
,
National Tsing Hua University
Siew Hong Teoh
,
University of California-Los Angeles
Jiawen Yan
,
Cornell University

Abstract

We investigate how perceived CEO trustworthiness shapes corporate innovation under uncertainty and incomplete contracting. Drawing on the well-documented psychological evidence linking facial features to trait perception, we use a machine-learning model to construct a unique face-based measure of perceived CEO trustworthiness. This measure correlates strongly with higher employee ratings, greater shareholder support in proxy votes, and faster analyst revisions of earnings forecasts. Furthermore, firms led by more trustworthy-looking CEOs generate both a larger number of and higher-quality patents. A causal interpretation is supported by a difference-in-differences-in-differences test based on local trust crises triggered by financial fraud of geographically proximate but economically unrelated firms. Further mechanism tests suggest that trustworthy CEOs can pursue riskier projects and are also more efficient in innovation. Further evidence suggests that CEO perceived trustworthiness reduces perceived information asymmetries across stakeholders. It is associated with lower analyst forecast dispersion, a stronger sensitivity of innovation output to employee stock options, and stronger support for R&D from short-term institutional investors. Taken together, our evidence highlights the role of trust as a form of social capital that facilitates high-risk, high-return innovation.
JEL Classifications
  • C6 - Mathematical Methods; Programming Models; Mathematical and Simulation Modeling