68 Program for Research in Markets and Organizations Generative Gender Gaps: Evidence from LLM Summarization

Authors:

Sarva Anbarasu, Dafna Bearson, Rembrand Koning, Tarun Khanna

Date Created:

2025-01-01

Course Title:

Professor:

Not specified

About Paper:

This ongoing research investigates whether gender differences evaluate gender differences in how users rate our AI-generated in AI preferences can reinforce gender inequalities in AI usage news summaries. Then, we will fine-tune the summarization tool during the fine-tuning process. We hypothesize that if AI based on majority male-preference data and assess whether user systems are disproportionately trained on data reflecting male ratingsshiftacrossgenders. Findingsfromthisstudyaimtoinform user preferences, the resulting outputs may marginalize other the design of more gender-inclusive AI systems and responsible groups and perpetuate bias. To explore this, we developed an fine-tuning practices. More broadly, our results could inspire AI news summarization tool that generates daily business and further research on demographic representation in AI systems and technology digests using generative AI. In the first phase of our their effectiveness across diverse user groups, which is important study, we administered surveys to assess gender differences in as AI becomes increasingly ubiquitous. general AI usage and attitudes. In the next phase, we will The Ownership Project M Aditta Arian, Tony Guidotti, Nien-he Hsieh Williams College | Economics | 2026 Despite extensive land tenure reforms across Sub-Saharan Africa productivity in Sub-Saharan Africa? (2) Do different types of over the past three decades, systematic evidence on their economic agricultural investments (soil conservation, tree planting, land impacts remains fragmented across individual country studies. fallowing) show different levels of responsiveness to improved This meta-analysis will synthesize quantitative evidence from tenure security? 50+ rigorous studies (1995-2025) examining how different tenure Our research will provide the most comprehensive quantitative systems affect agricultural productivity, investment behavior, and synthesis of African land tenure impacts. While we acknowledge in turn poverty reduction across the continent. the complexity of poverty pathways and contextual variations, our We will analyze standardized effect sizes from randomized analysis will focus on establishing robust average effect sizes for controlled trials, quasi-experimental designs, and panel studies key economic outcomes. These findings will deepen evidence- covering multiple African countries. Using random-effects meta- based understanding of land-tenure reforms in the African context analysis, we will examine two focused research questions: (1) and lay the groundwork for future studies that explore the What is the average effect of land tenure security on agricultural transferability of these insights to other developing regions. Innovation Engines: How Legendary Corporate Labs Shaped the 20th Century Alex Banull, Kyle Myers Brown University | Economics | 2027 Although corporate laboratories are the foundation of sustained the research group as an idea develops, promoting risk-taking economic development, the uncertain nature of research makes behavior, and calculating the expected economic valuation of the them difficult to structure. Innovations are relatively unpredictableea to determine if it should be terminated. Most notably, Bell with a low probability of discovery, and the development of an Labs primarily followed these theoretical management strategies idea is also highly uncertain. This variability poses a problem forfor innovation, while Xerox PARC disregarded them. Although measuring the effectiveness of a laboratory, because in the short both laboratories had access to ample funding, infrastructure, and run, output is not necessarily indicative of a successful strategy.skilledresearchers, thisdifferenceinmanagementisafundamental Due to this uncertainty, many laboratories throughout history have reason for Bell Labs’ economic success and Xerox PARC’s overlooked and underdeveloped certain innovations. We analyzed failure. ThecomparisonbetweenBellandPARCiscurrentlybeing foundational economic papers within innovation and then applied developed into a case study to further understand their difference them to historical examples of corporate laboratories to find time-in outcome. We are also working on finding more records on the invariant strategies that incentivize innovation and development. compensation and insurance structure of management. Effective structures include gradually restricting the autonomy of Corporate Engagement with Societal Issues in the U.S. Evelyn Cheng, Magdalena Larreboure, Vincent Pons University of California Berkeley | Economics | 2026 This project examines how S&P 500 companies engaged with �So far, we have found that 392 out of 500 companies (78.4%) polarizing sociopolitical issues in the United States between 2008 publicly addressed at least one issue, with climate change alone and 2025. We built a dataset of nearly 40,000 press releases accounting for 67% of all stance-taking events. Only 7% and news articles from companies that were part of the S&P 500 of stances were issued by CEOs, who more often spoke on during this period, focusing on nine partisan topics: abortion, gun violence, immigration, and minimum wage. Preliminary climatechange, genderequality, gunviolence, immigrationpolicy, evidence suggests that workforce composition partially influences LGBTQ issues, minimum wage, race discrimination, and voting engagement, though effects vary by topic. laws. Looking ahead, we aim to explore whether corporate ideology �We organized our analysis around six key areas: (i) which shapes the timing of engagement—for example, whether liberal- companies take stances, (ii) what types of stances they take, (iii)leaning firms are more vocal under Democratic administrations. how the demographic and political characteristics of employees We also plan to analyze spatial and temporal patterns in stance- and leadership correlate with the number and type of stances taking, as well as connections between traditional political activity taken, (iv)howotherformsofcorporatepoliticalengagement(e.g., and public issue engagement. political donations and lobbying) relate to stance-taking, (v) when companies take stances, and (vi) where such stances are taken. 70 Program for Research in Markets and Organizations A New (and Very Old) View of Tax Policy: Benefit-Based Taxation Esther Cho, Jo Ellery, Matthew Weinzierl Harvard College | Quincy House | Economics | 2027 Income redistribution aims to reduce inequality and increase Clearly, global redistribution does not follow a utilitarian welfare within a utilitarian framework. Within a utilitarian framework, but principles for optimal policy such as benefit-based framework, a global approach to redistribute, which transfers taxation, predict that we should instead see a continuation in money across countries, produces maximal levels of utility in the the standard logic where redistribution should increase when we context of societal welfare. This is largely due to the fact that move from the national to the global level, revealing a potential wealth inequality levels continue to grow as we consider larger mismatch. We work to unravel the puzzle of redistribution’s jurisdiction sizes, when comparing local, state, federal, and global inverted-u relationship to jurisdictional scale through quantitative levels. Thus, utility would disproportionately increase with global analysis on the redistribution rates at each level of government, redistribution in comparison to local or domestic redistribution. along with exploring or modifying various theoretical frameworks that may work to explain this pattern where standard utilitarianism However, in practice, redistribution follows a different pattern, fails. This may have future implications for global redistribution increasing nearly exponentially from the local government to state government to the federal government but then facing a sharp and the adoption of a benefit-based approach. downturn when we consider international aid. Estimating Legislative and Regulatory Risks Caleb Eynon, Max Miller University of Alabama | Economics and Mathematics | 2026 Firms that operate in the USA must do so within the boundaries of we use the relationship between daily company-level stock returns current legislation. For this reason, they, and firms they interact and changes in the probability of passage to create a measure of with, are exposed to legislative risk—risks associated with present exposure to each bill, (e.g. companies with a larger regression and future legislation. This paper provides a flexible methodology coefficient are more exposed to that legislation). As this project to measure exposure to this risk for publicly-traded companies. progresses, we will provide descriptive statistics on exposure We do this in two steps. First, we use data on legislation from acrossindustriesandfirms. Thisdatacanthenbeusedinanalyzing 2003-2021 to predict the probability that each individual bill will firm-levelactivity,suchaslobbyingorinvestmentdecisions. More become law. This is accomplished with a recurrent neural network broadly, this paper makes advances at the intersection of politics (RNN) which can recognize patterns in sequential data. Second, and finance. Sanctioned Radicals: Authorization and Discipline in Black and Asian Employee Resource Groups Jaycelin Eyre, Jaylon Sherrell, Lumumba Seegars Brigham Young University | Psychology | 2026 How do marginalized groups collectively organize around racial tensions of sanctioning: balancing organizational legitimacy with identities in the workplace? The formation of employee resource the need to avoid disciplinary action. The Black ERG navigated groups (ERGs) within organizations comes with the double-sided dynamics by building support with Black executives and creating sword ofsanctioning. Sanctioningcanmeaneitherformalapproval a private chat room to discuss Black experiences, while presenting or punishment through, respectively, provision of resources or themselves guardedly in supervised settings. Asian ERG members disciplining. ERGs benefit from organizational authorization experienced salient differences between East/South Asian and but must navigate racialized hierarchies. This includes not just immigrant/multigenerational identities and events. However, they restricted access to resources, but systemic disciplining that shapesited under an Americanized Asian identity recognized by the how members behave and are perceived within organizations. company and shared events to make culture seem less “foreign” to PreviousresearchsuggestsBlackandAsiangroupsmightapproach colleagues. These behaviors permitted both ERGs authorization to sanctioning differently. To understand how collective organizing pursue goals while avoiding disciplining. These findings highlight around race manifested similarly and differently for Black versus how racialization constrains access to resources and reproduces Asian ERGs, we independently coded interviews with 31 Black marginalization through group-specific sanctioning dynamics. By ERG and 20 Asian ERG members from a major American-based showing that Black and Asian ERGs faced qualitatively different tech company. Coding themes from prior analysis were refined forms of racialization, this research points to the need for future iteratively, with the final iteration comparing ERG behaviors, studies to further examine how the nature of racialization shapes resources, and concerns. We found that, although surface-level collective organization across organizational contexts. behaviors appeared dissimilar, both groups navigated underlying

Abstract:

This ongoing research investigates whether gender differences evaluate gender differences in how users rate our AI-generated in AI preferences can reinforce gender inequalities in AI usage news summaries. Then, we will fine-tune the summarization tool during the fine-tuning process. We hypothesize that if AI based on majority male-preference data and assess whether user systems are disproportionately trained on data reflecting male ratingsshiftacrossgenders. Findingsfromthisstudyaimtoinform user preferences, the resulting outputs may marginalize other the design of more gender-inclusive AI systems and responsible groups and perpetuate bias. To explore this, we developed an fine-tuning practices. More broadly, our results could inspire AI news summarization tool that generates daily business and further research on demographic representation in AI systems and technology digests using generative AI. In the first phase of our their effectiveness across diverse user groups, which is important study, we administered surveys to assess gender differences in as AI becomes increasingly ubiquitous. general AI usage and attitudes. In the next phase, we will The Ownership Project M Aditta Arian, Tony Guidotti, Nien-he Hsieh Williams College | Economics | 2026 Despite extensive land tenure reforms across Sub-Saharan Africa productivity in Sub-Saharan Africa? (2) Do different types of over the past three decades, systematic evidence on their economic agricultural investments (soil conservation, tree planting, land impacts remains fragmented across individual country studies. fallowing) show different levels of responsiveness to improved This meta-analysis will synthesize quantitative evidence from tenure security? 50+ rigorous studies (1995-2025) examining how different tenure Our research will provide the most comprehensive quantitative systems affect agricultural productivity, investment behavior, and synthesis of African land tenure impacts. While we acknowledge in turn poverty reduction across the continent. the complexity of poverty pathways and contextual variations, our We will analyze standardized effect sizes from randomized analysis will focus on establishing robust average effect sizes for controlled trials, quasi-experimental designs, and panel studies key economic outcomes. These findings will deepen evidence- covering multiple African countries. Using random-effects meta- based understanding of land-tenure reforms in the African context analysis, we will examine two focused research questions: (1) and lay the groundwork for future studies that explore the What is the average effect of land tenure security on agricultural transferability of these insights to other developing regions. Innovation Engines: How Legendary Corporate Labs Shaped the 20th Century Alex Banull, Kyle Myers Brown University | Economics | 2027 Although corporate laboratories are the foundation of sustained the research group as an idea develops, promoting risk-taking economic development, the uncertain nature of research makes behavior, and calculating the expected economic valuation of the them difficult to structure. Innovations are relatively unpredictableea to determine if it should be terminated. Most notably, Bell with a low probability of discovery, and the development of an Labs primarily followed these theoretical management strategies idea is also highly uncertain. This variability poses a problem forfor innovation, while Xerox PARC disregarded them. Although measuring the effectiveness of a laboratory, because in the short both laboratories had access to ample funding, infrastructure, and run, output is not necessarily indicative of a successful strategy.skilledresearchers, thisdifferenceinmanagementisafundamental Due to this uncertainty, many laboratories throughout history have reason for Bell Labs’ economic success and Xerox PARC’s overlooked and underdeveloped certain innovations. We analyzed failure. ThecomparisonbetweenBellandPARCiscurrentlybeing foundational economic papers within innovation and then applied developed into a case study to further understand their difference them to historical examples of corporate laboratories to find time-in outcome. We are also working on finding more records on the invariant strategies that incentivize innovation and development. compensation and insurance structure of management. Effective structures include gradually restricting the autonomy of Corporate Engagement with Societal Issues in the U.S. Evelyn Cheng, Magdalena Larreboure, Vincent Pons University of California Berkeley | Economics | 2026 This project examines how S&P 500 companies engaged with �So far, we have found that 392 out of 500 companies (78.4%) polarizing sociopolitical issues in the United States between 2008 publicly addressed at least one issue, with climate change alone and 2025. We built a dataset of nearly 40,000 press releases accounting for 67% of all stance-taking events. Only 7% and news articles from companies that were part of the S&P 500 of stances were issued by CEOs, who more often spoke on during this period, focusing on nine partisan topics: abortion, gun violence, immigration, and minimum wage. Preliminary climatechange, genderequality, gunviolence, immigrationpolicy, evidence suggests that workforce composition partially influences LGBTQ issues, minimum wage, race discrimination, and voting engagement, though effects vary by topic. laws. Looking ahead, we aim to explore whether corporate ideology �We organized our analysis around six key areas: (i) which shapes the timing of engagement—for example, whether liberal- companies take stances, (ii) what types of stances they take, (iii)leaning firms are more vocal under Democratic administrations. how the demographic and political characteristics of employees We also plan to analyze spatial and temporal patterns in stance- and leadership correlate with the number and type of stances taking, as well as connections between traditional political activity taken, (iv)howotherformsofcorporatepoliticalengagement(e.g., and public issue engagement. political donations and lobbying) relate to stance-taking, (v) when companies take stances, and (vi) where such stances are taken. 70 Program for Research in Markets and Organizations A New (and Very Old) View of Tax Policy: Benefit-Based Taxation Esther Cho, Jo Ellery, Matthew Weinzierl Harvard College | Quincy House | Economics | 2027 Income redistribution aims to reduce inequality and increase Clearly, global redistribution does not follow a utilitarian welfare within a utilitarian framework. Within a utilitarian framework, but principles for optimal policy such as benefit-based framework, a global approach to redistribute, which transfers taxation, predict that we should instead see a continuation in money across countries, produces maximal levels of utility in the the standard logic where redistribution should increase when we context of societal welfare. This is largely due to the fact that move from the national to the global level, revealing a potential wealth inequality levels continue to grow as we consider larger mismatch. We work to unravel the puzzle of redistribution’s jurisdiction sizes, when comparing local, state, federal, and global inverted-u relationship to jurisdictional scale through quantitative levels. Thus, utility would disproportionately increase with global analysis on the redistribution rates at each level of government, redistribution in comparison to local or domestic redistribution. along with exploring or modifying various theoretical frameworks that may work to explain this pattern where standard utilitarianism However, in practice, redistribution follows a different pattern, fails. This may have future implications for global redistribution increasing nearly exponentially from the local government to state government to the federal government but then facing a sharp and the adoption of a benefit-based approach. downturn when we consider international aid. Estimating Legislative and Regulatory Risks Caleb Eynon, Max Miller University of Alabama | Economics and Mathematics | 2026 Firms that operate in the USA must do so within the boundaries of we use the relationship between daily company-level stock returns current legislation. For this reason, they, and firms they interact and changes in the probability of passage to create a measure of with, are exposed to legislative risk—risks associated with present exposure to each bill, (e.g. companies with a larger regression and future legislation. This paper provides a flexible methodology coefficient are more exposed to that legislation). As this project to measure exposure to this risk for publicly-traded companies. progresses, we will provide descriptive statistics on exposure We do this in two steps. First, we use data on legislation from acrossindustriesandfirms. Thisdatacanthenbeusedinanalyzing 2003-2021 to predict the probability that each individual bill will firm-levelactivity,suchaslobbyingorinvestmentdecisions. More become law. This is accomplished with a recurrent neural network broadly, this paper makes advances at the intersection of politics (RNN) which can recognize patterns in sequential data. Second, and finance. Sanctioned Radicals: Authorization and Discipline in Black and Asian Employee Resource Groups Jaycelin Eyre, Jaylon Sherrell, Lumumba Seegars Brigham Young University | Psychology | 2026 How do marginalized groups collectively organize around racial tensions of sanctioning: balancing organizational legitimacy with identities in the workplace? The formation of employee resource the need to avoid disciplinary action. The Black ERG navigated groups (ERGs) within organizations comes with the double-sided dynamics by building support with Black executives and creating sword ofsanctioning. Sanctioningcanmeaneitherformalapproval a private chat room to discuss Black experiences, while presenting or punishment through, respectively, provision of resources or themselves guardedly in supervised settings. Asian ERG members disciplining. ERGs benefit from organizational authorization experienced salient differences between East/South Asian and but must navigate racialized hierarchies. This includes not just immigrant/multigenerational identities and events. However, they restricted access to resources, but systemic disciplining that shapesited under an Americanized Asian identity recognized by the how members behave and are perceived within organizations. company and shared events to make culture seem less “foreign” to PreviousresearchsuggestsBlackandAsiangroupsmightapproach colleagues. These behaviors permitted both ERGs authorization to sanctioning differently. To understand how collective organizing pursue goals while avoiding disciplining. These findings highlight around race manifested similarly and differently for Black versus how racialization constrains access to resources and reproduces Asian ERGs, we independently coded interviews with 31 Black marginalization through group-specific sanctioning dynamics. By ERG and 20 Asian ERG members from a major American-based showing that Black and Asian ERGs faced qualitatively different tech company. Coding themes from prior analysis were refined forms of racialization, this research points to the need for future iteratively, with the final iteration comparing ERG behaviors, studies to further examine how the nature of racialization shapes resources, and concerns. We found that, although surface-level collective organization across organizational contexts. behaviors appeared dissimilar, both groups navigated underlying

Source:

Harvard / Harvard College | Winthrop House | Computer Science | 2027 / 2025

Topics:

economic, company, stance, redistribution, group, black, asian, gender, acros, level, innovation, global

Co-authors:

@sarvaanbarasu210 , @dafnabearson211 , @rembrandkoning212 , @tarunkhanna213