Research

Study Overview

Depression and anxiety have risen sharply among U.S. college students, yet a large share of affected students receive no clinical care. A central barrier is navigation: students encounter fragmented/static resources with no personalized front-door, no adaptive-triage, and limited support before distress escalates. This study evaluates Flourish@Berkeley, a purpose-built AI wellness platform embedded in UC Berkeley’s University Health Services (UHS), in the context of an at-scale randomized controlled trial (RCT) with 10,000 students. Treatment-assigned students gain access to an app with an AI companion, which delivers evidence-based micro-practices for coping, emotion regulation, problem solving, and positive functioning, scaffolding them into durable habits, and dynamically. routes students with conversational navigation to UC Berkeley’s campus resource ecosystem (e.g., financial aid, academic and peer support, wellness, UHS resources, crisis and clinical referrals). Students assigned to the control group retain full access to existing resources via traditional channels. The study tests whether low-stigma, institutionally integrated AI support reduces unmet need, improves validated survey measures of depression, anxiety, well-being, and cognitive functioning. We will also measure the impact on administrative measures of academic progress, service utilization (i.e., changes in who reaches which types of care) and longer-run labor market outcomes. The zero-infrastructure solution minimizes adoption frictions; the deployment model would be directly scalable across the UC system and other U.S. colleges, including community colleges with more limited services/infrastructure.

Study Results

Pending

Populations: University students

News & media

AI Doesn’t Reduce Work—It Intensifies It

February 9, 2026

Right now, many companies are worried about how to get more employees to use AI. After all, the promise of AI reducing the burden of some work—drafting routine documents, summarizing information, and debugging code—and allowing workers more time for high-value tasks is tantalizing.

AI promised to free up workers’ time. UC Berkeley Haas researchers found the opposite.

February 18, 2026

While conducting research on how AI was changing daily work at a U.S. technology company, UC Berkeley Haas doctoral student Xingqi Maggie Ye noticed a pattern that raised a provocative question: What if AI is intensifying work rather than reducing it?

AI Doesn’t Reduce Work—It Intensifies It

March 3, 2026

One of the promises of AI is that it can reduce workloads so employees can focus more on higher-value and more engaging tasks. But according to new research, AI tools don’t reduce work, they consistently intensify it.

Study Overview

Whether generative AI narrows or widens inequality in low-income economies, and whether it builds or erodes human capital, remains largely untested outside high-income, short-horizon settings. We study these questions among young entrepreneurs in Uganda by embedding a randomized take-up experiment in the SEED 2.0 program, which in 2025 randomized 8,455 secondary-school graduates into a control group and two entrepreneurship-training arms. At the 2027 follow-up, we offer a purpose-built AI assistant—trained on the program curriculum and locally developed content and delivered over WhatsApp—to three quarters of participants in each arm, retaining the rest as a within-arm comparison group, with assignment stratified by baseline knowledge, prior training take-up, and gender. Among those offered, we cross-randomize the tool's mode (tutoring versus direct answers) and a take-up incentive. Intention-to-treat comparisons identify the effect of access and how it varies with baseline ability—whether AI complements or substitutes for skill—whether learning persists once the tool is withdrawn, whether adoption is itself selected on ability, and whether AI substitutes for or complements training. The study tests not only whether AI helps, but whether its design and deployment can make the benefits inclusive.

Study Overview

This project examines whether technology-enabled enforcement can strengthen legal protections for women at risk of domestic violence. Conducted in partnership with Brazil’s Ministry of Justice, the study will evaluate Brazil’s new bilateral electronic monitoring program, which combines GPS ankle bracelets for perpetrators with real-time alerts to victims when court-ordered protection perimeters are breached. Brazil has a strong legal framework for domestic violence, but enforcement remains limited. In 2025, femicides reached a record high, with 1,568 women killed, more than 945,000 new domestic violence cases processed, and over 619,000 protective orders granted. At least 18.3% of protective orders are violated, and 13.1% of women killed had an active protective order. The research will study whether bilateral monitoring improves compliance with protective orders, accelerates enforcement, deters further violence, and strengthens victims’ safety, mental health, labor market participation, and agency.

Study Overview

We study the impacts of an entrepreneurship business training on negotiation skills. The Skills for Effective Entrepreneurship Development (SEED) program is 3-week mini-MBA for youth who just graduated secondary school in Uganda. SEED was implemented as an at scale Randomized Controlled Trial (RCT) and presented two separate curricula. The hard skills MBA featured a mix of approximately 75% hard-skills and 25% soft-skills; the soft-skills curriculum had the reverse mix. Only youth in the Soft SEED arm received explicit training in communication, negotiation, and persuasion. Four years after the intervention, we conducted a lab-in-the-field experiment and collected speech-to-text data as youth engaged in incentivized negotiation involving financing and start dates for a possible sale agreement. Study participants in the soft SEED group struck better deals than the control and hard SEED groups. Leveraging large language model (LLM) annotation, we investigate how negotiation strategies profiles differ across experimental groups. Preliminary results suggest that youth in the soft SEED group adopted more aggressive stances in their initial offers; their communication style was more assertive, but they were also more able to manage disagreement, avoiding communication breakdowns. Future work will provide richer characterization of study participants' negotiation profiles by examining win-win strategies, which were an explicit focus of the soft SEED training.

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

Study Overview

This project evaluates the scale-up of Skills for Effective Entrepreneurship Development (SEED) through a randomized controlled trial involving more than 8,000 Ugandan youth. SEED is an intensive three-week residential “mini-MBA” that combines business, entrepreneurship, and broader skills training and was first evaluated experimentally in Uganda in 2013. The new study will test whether the program retains its effectiveness when implemented at scale and estimate the incremental value of supplementing the core curriculum with a digital business skills module and a generative AI tool. The SEED-tailored, AI-powered virtual tutor will support participants during training and remain available as they launch and develop their businesses. The trial will generate rigorous evidence on the scalability of a previously tested youth entrepreneurship program and on whether sustained access to AI-enabled guidance can strengthen business practices, productivity, entrepreneurship, and economic opportunities among young people.

Populations: Small and medium enterprises (SME)

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

Study Overview

Women- and youth-led SMEs in East Africa face complementary constraints to growth — limited business capabilities, restricted access to finance, and indivisible production infrastructure. Which of these bind, and for which entrepreneurs, remains an open empirical question. We propose a pilot in Kenya with Somo Africa to lay the foundations for a full-scale RCT of Somo's integrated training–financing–hub model, in which financing decisions are driven by the Somo Scorecard, a proprietary AI-based alternative-data credit score that replaces collateral and formal credit history with performance data derived from Somo's DigiKua record-keeping tool. The pilot is designed to (i) validate a two-stage randomization against Somo's programmatic funnel; (ii) quantify take-up, compliance, and attrition by arm and by Somo Scorecard tier — in particular, whether low financing take-up reflects selection of not-yet-finance-ready applicants, an AI scorecard calibrated too tightly, or demand-side risk aversion; (iii) characterize outcome distributions needed for power calculations for the full RCT; (iv) validate instruments including DigiKua-derived business records that serve as the alternative data underlying the Scorecard; (v) identify operational considerations to resolve before scale-up. These learnings will shape the full RCT that can identify which constraints bind and for whom — and whether AI-based alternative-data credit scoring can independently identify the women entrepreneurs for whom credit is most productive.

Intervention Partner: Somo

Populations: Small, micro enterprises, Small businesses

Study Overview

This project evaluates whether artificial intelligence (AI) can improve competition, efficiency, and inclusion in public procurement. In partnership with ChileCompra and DIPRES, the research will develop and rigorously test two AI tools: one that helps public buyers draft clearer, standardized procurement specifications, and another that helps suppliers—particularly small and medium enterprises (SMEs)—identify relevant opportunities and prepare more competitive bids. Using a large-scale randomized controlled trial on Chile's national procurement platform, the study will measure impacts on bidder participation, procurement prices, SME success, and government spending efficiency. The project aims to generate evidence on how AI can strengthen public institutions, expand economic opportunity, and improve the delivery of public services through more effective and competitive procurement systems.

Intervention Partner: ChileCompra

Populations: Small and medium enterprises (SME)

News & media

ChileCompra, University of Chile, and UC Berkeley expand collaboration to apply AI in public procurement

December 15, 2025

The new project will use advanced artificial intelligence models to improve the efficiency, integrity, and participation of SMEs in the government procurement system, with funding from Schmidt Sciences.

Study Overview

In growth markets, where an estimated 1.4 billion individuals remain unbanked, the reliable and consistent delivery of last-mile banking services remains a critical challenge despite multi-sector technological advancements. In emerging markets, neighborhood mom-and-pop store owners acting as micro-fintech agents are crucial basic banking service providers in underserved communities. However, efforts to improve access to banking services through these agents often face a significant challenge known as 'merchant dormancy'. Empowering last-mile micro-fintech agents with meaningful, actionable information on financial products and how to manage and maintain them could reduce dormancy, potentially catalyzing an estimated USD 380 billion in annual economic value and fostering local economic development. For micro-fintech agents serving the underbanked in emerging markets, the AI revolution offers an opportunity to jumpstart their productivity supported by access to personalized and timely advice. This research leverages Large Language Models (LLMs) in two ways a) Personalized recommendation (curation) with an AI virtual secretary and b) Low-cost and fast content creation through Generative AI models aimed to reduce merchant dormancy, increase access to and improve quality of financial services offered by micro-fintech agents in rural and peri-urban India.

Study Results

Pending

Intervention: AI-generated personal support for micro-fintech agents

Research Partner: Center for Growth Markets

Intervention Partner: FINO

Populations: Small, micro enterprises, unbanked households and individuals

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

News & media

Revolutionizing Financial Inclusion:AI-Powered Personalized Support for Last-Mile Banking

September 1, 2024

We’re pioneering an innovative approach that leverages the power of Large Language Models (LLMs) to democratize content personalization for small fintech merchant support, bridging the data gap that has long hindered effective service delivery.

Empowering financial inclusion in India through reliable micro-fintech agents

September 1, 2024

We have developed an AI-powered solution that offers personalized support to micro-FinTech agents in data-scarce environments. By leveraging LLMs and Generative AI, we deliver timely and relevant information that empowers agents to thrive in their roles.

Study Overview

Algorithmic lending has led to a rise in personalized pricing reflecting information beyond the credit score. We study the welfare effects of pricing on increasingly granular information. We use a randomized pricing experiment and causal machine learning to estimate demand and cost curves and embed them in competitive equilibrium. Finer pricing need not improve welfare: relative to uniform pricing, credit-score pricing reduces total surplus, as prices increase in adversely selected low-score pools, pushing out safest borrowers.

Study Results

Personalization improves on the score by separating observables predicting risk levels from those predicting selection, but uniform pricing dominates, as lending gains remain modest.

Intervention: Randomized variations in interest rates for consumer loans

Populations: Middle income households

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

Study Overview

Organizations learn over time. They build organizational capital and form beliefs about their fundamentals. Motivated by recent advances in AI, we study organizational learning technologies that accelerate both processes.

Study Results

We show that, in a large class of models of firm dynamics, the value of organizational learning technologies (VOLT) is governed by two simple statistics: the relative size and lifespan of mature firms. In the United States, VOLT is on the order of one GDP — implying that organizational learning technologies like AI have the potential to double aggregate output. Much of VOLT reflects increases in average firm lifespans rather than productivity. Across industries, VOLT varies widely and is orthogonal to existing AI exposure measures. Overall, our results point to faster organizational learning as a meaningful and distinct channel of AI’s transformative potential, beyond production automation and scientific discovery.

Working Paper: NBER Conference Working Paper

News & media

A new measure finds AI could double U.S. economic output by helping businesses learn faster—or fail faster

April 10, 2026

New UC Berkeley Haas research finds that AI-driven organizational learning could roughly double U.S. economic output in the long run, based on a new metric called VOLT (Value of Organizational Learning Technologies).

Study Overview

Scalable and low-cost artificial intelligence (AI) assistance has the potential to improve firm decision making and economic performance, particularly in emerging markets. However, running a business involves a wide range of open-ended problems, making it unclear whether and how recent advances in AI can help business owners around the world make better decisions. In a field experiment with Kenyan entrepreneurs, we evaluated the impact of AI advice on small business revenues and profits by randomizing access to a GPT-4-powered AI business assistant.

Study Results

Although we are unable to reject the null hypothesis of no average treatment effect on firm revenues and profits, we find that the effect for entrepreneurs who were low performing at baseline is over 0.20-standard-deviations lower than for initial high performers. Subsample analyses show that low performers did nearly 10% worse because of the AI assistant, whereas high performers may have benefited by over 15%. This differential impact does not appear to result from differences in the questions posed to the AI or the advice that it provided but rather, from the advice that entrepreneurs chose to implement. More broadly, these results show that generative AI is already capable of impacting real-world business performance—although in uneven and sometimes unexpected ways.

Journal Publication: Nicholas G. Otis, Rowan Clarke, Solène Delecourt, David Holtz, Rembrand Koning (2026) The Uneven Impact of Generative Artificial Intelligence on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya. Management Science 0(0). https://doi.org/10.1287/mnsc.2024.06909

News & media

Gen AI field experiment shows mixed results in helping small businesses grow

April 1, 2024

While generative AI may hold promise as an efficient way to help small businesses grow, a study of entrepreneurs in Kenya found real-world limitations for those businesses that need it most.

How AI Helps the Best and Hurts the Rest

April 20, 2026

Generative AI can boost performance for stronger business owners but harm those already struggling. The difference comes down to human judgment.

Study Overview

We investigate the calibration of large language models' (LLMs') confidence across diverse tasks. The results of our preregistered study show that the current crop of LLMs are, like people, too sure they are right: confidence exceeds accuracy, on average. Importantly, however, this tendency is moderated by a powerful hard-easy effect, wherein overconfidence is greatest on difficult tests; by contrast, easy tests actually show substantial underconfidence. We develop LifeEval, a test for evaluating model calibration across levels of difficulty.

Working Paper: Michael, N., BenShushan, D., Bien, J., & Moore, D. A. (2026). Confidence calibration in large language models. arXiv preprint arXiv:2605.23909. https://doi.org/10.48550/arXiv.2605.23909

Study Overview

We study how a frontier AI model affects scientific discovery by examining the release of the AlphaFold2 algorithm and its impact on structural biology and related fields of science. Structural biology is the field of science concerned with understanding the structure and function of proteins. Researchers in this field historically devoted substantial time and resources to experimentally solving three-dimensional protein structures. AlphaFold can predict these structures without running experiments.

Study Results

In July 2021, researchers gained access to hundreds of thousands of these AI-predicted structures virtually overnight. Yet, to date, we find that the rate of experimental structure determination has remained almost unchanged. Instead, researchers appear to use predicted structures to facilitate and complement experimental structure determination. Looking at downstream science that builds on protein structures, we find that basic research on proteins that had no structure information prior to AlphaFold increases by 15 to 40% relative to proteins that already had a structure, shifting the direction of research toward less-studied proteins. However, we find no evidence so far that more applied, early-stage drug development is targeting these proteins, though such activity may emerge in the future.

Working Paper: Ryan R. Hill and Carolyn Stein, "How Artificial Intelligence Shapes Science: Evidence from AlphaFold," NBER Working Paper 35143 (2026), https://doi.org/10.3386/w35143.

News & media

AI is expanding the boundaries of biological research. Will drug development follow?

May 27, 2026

A scant few years after a powerful artificial intelligence tool was hailed as a revolution that would transform drug discovery, a new UC Berkeley Haas study finds that Google’s AlphaFold2 has fundamentally broadened the landscape of scientific research by unlocking the structure of millions of proteins that had been too poorly understood to study.

Study Overview

While prior research has examined the consequences of corporate environmental, social, and governance (ESG) activities, less is known about firms’ underlying motives for engaging in ESG. We use artificial intelligence to analyze annual ESG reports and classify firms’ ESG motivations into three categories: moral duty or altruism (ALTRUISM), value creation or market demand (MARKET), and regulatory compliance or risk management (RISK). We construct firm-year measures capturing the relative prominence of these underlying motivational logics and document substantial cross-sectional and temporal variation.

Study Results

We find that ESG motivation is systematically related to firms’ fundamentals, organizational environments, and governance characteristics, and that motivations shift around internal shocks and salient external events. Firms emphasizing strategic motives—market and risk rationales—are associated with higher ESG performance relative to firms emphasizing altruistic motives. Mediation analyses suggest that this association operates in part through deeper integration of ESG into broader operational strategy and governance structures. We further show that ESG motivation is related to the types of ESG activities firms pursue: risk-motivated firms exhibit fewer ESG controversies and violations, while altruism-oriented firms engage in more charitable donations. Overall, our findings highlight how differences in firms’ underlying ESG motivations are associated with variation in ESG implementation and outcomes, underscoring the importance of considering the “why” behind corporate ESG engagement.

Working Paper: Kim, Shawn and Kim, Sun Tae and Korri, Piia, Does Why Matter? Motivation behind Corporate ESG Activities (March 14, 2026). Available at SSRN: https://ssrn.com/abstract=5525359 or http://dx.doi.org/10.2139/ssrn.5525359

Study Overview

Haraka leverages blockchain, local stablecoins, and community trust to reimagine micro-finance and expand financial inclusion and health of underserved populations. Savings circles are a crucial segment of last-mile financial services for the estimated 510 million individuals worldwide who have turned to informal and self-organized village savings and loan associations (VSLAs). Despite the crucial role that VSLAs play in providing services to underbanked people, especially women, savings groups remain excluded from access to formal credit. The goal of this research is to understand how access to credit for savings groups and their members might be expanded by coupling machine learning methods with data on VSLA transactions, digital records, and information on the groups’ social capital, reputation and cohesion when assessing creditworthiness. Technology is rapidly digitizing many aspects of saving groups’ functioning, including group interactions, access to information, digital record-keeping, and electronic transactions creating invaluable digital footprints. This information is then combined to build  a self-sovereign AI/ML credit score for individuals, enabling them to access financial opportunities beyond their immediate communities. Beyond traditional measures of access to credit and creditworthiness, we are also interested in possible impacts on members’ economic activities and their performance, as well as downstream impacts on wellbeing and gender empowerment.

Study Results

This study ended after Haraka closed in 2026.

Intervention: Innovative credit scoring to expand access to financial services leveraging blockchain technology

Intervention Partner: Haraka

Populations: Rural, underbanked in community savings and loan associations

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

News & media

Leveraging reputation and local stablecoins to improve credit access for unbanked VSLAs in Ghana

December 3, 2024

Mercy Corps Ventures, Haraka, and Grameen Foundation launch a new pilot to increase financial access for unbanked women, testing the use of social reputation and social guarantees to underwrite DeFi microloans issued in local currency stablecoins in Ghana.

Study Overview

Over the next two decades, millions of small business owners will face retirement, yet the majority are unprepared–only one-third have an exit plan in place. While numerous buyers are available for many small business owners, selling to employees offers an opportunity to ensure continuity while (usually) enjoying tax advantages. However, despite its benefits, most business owners are unaware of employee ownership as a viable exit strategy. Even for those interested, determining whether an employee ownership conversion is the right fit for their business can be complex and uncertain. Employee ownership conversions require specialized expertise that most business lawyers, lenders, and advisors do not have. This report describes a chatbot designed to help business owners learn if worker ownership might be a good fit for their business.

Study Results

We find that AI chatbots can be a useful tool to advance worker ownership. First, it lowers the barriers for potential sellers to learn about selling to employees. Second, it automates the initial steps of the conversion process for those who advise on worker ownership, freeing up their time for deals likely to go through. We conclude by discussing the implications of this finding and suggestions for future research and chatbot development. While this version focuses only on selling to worker-owners, a natural extension can help any business owner identify plausible buyers such as family members, current managers, and outsiders.

Intervention: Employee ownership models

Research Partner: Institute for the Study of Employee Ownership and Profit Sharing

Populations: Low-wage workers

Working Paper: Foley, William and Drew McArthur. 2025. Article 9: Examining the Feasibility of a Worker-Ownership Conversion AI Chatbot

News & media

The Promote Ownership by Workers for Economic Recovery Act (AB 2849) Panel

June 13, 2023

The Promote Ownership by Workers for Economic Recovery Act (AB 2849), codified in Labor Code sections 10000-10010) establishes a panel to study the creation of an Association of Cooperative Labor Contractors, among other potential activities, to facilitate the growth of democratically-run high-road cooperative labor contractors.

Study Overview

Despite growing excitement (and concern) about the fast adoption of generative artificial intelligence (Gen AI) across all academic disciplines, empirical evidence remains fragmented, and systematic understanding of the impact of large language models (LLMs) across scientific domains is limited.

Study Results

We analyzed large-scale data from three major preprint repositories to show that the use of LLMs accelerates manuscript output, reduces barriers for non-native English speakers, and diversifies the discovery of prior literatures. However, traditional signals of scientific quality such as language complexity are becoming unreliable indicators of merit, just as we are experiencing an upswing in the quantity of scientific work. As AI systems advance, they will challenge our fundamental assumptions about research quality, scholarly communication, and the nature of intellectual labor. Science policy-makers must consider how to evolve our scientific institutions to accommodate the rapidly changing scientific production process.

Journal Publication: Keigo Kusumegi et al. ,Scientific production in the era of large language models.Science390,1240-1243(2025).DOI:10.1126/science.adw3000

Study Overview

Are widespread stereotypes accurate or socially distorted? This continuing debate is limited by the lack of large-scale multimodal data on stereotypical associations and the inability to compare these to ground truth indicators. Here we overcame these challenges in the analysis of age-related gender bias, for which age provides an objective anchor for evaluating stereotype accuracy.

Study Results

Despite there being no systematic age differences between women and men in the workforce according to the US Census, we found that women are represented as younger than men across occupations and social roles in nearly 1.4 million images and videos from Google, Wikipedia, IMDb, Flickr and YouTube, as well as in nine language models trained on billions of words from the internet. This age gap is the starkest for content depicting occupations with higher status and earnings. We demonstrate how mainstream algorithms amplify this bias. A nationally representative pre-registered experiment (n = 459) found that Googling images of occupations amplifies age-related gender bias in participants’ beliefs and hiring preferences. Furthermore, when generating and evaluating resumes, ChatGPT assumes that women are younger and less experienced, rating older male applicants as of higher quality. Our study shows how gender and age are jointly distorted throughout the internet and its mediating algorithms, thereby revealing critical challenges and opportunities in the fight against inequality.

Journal Publication: Guilbeault, D., Delecourt, S. & Desikan, B.S. Age and gender distortion in online media and large language models. Nature 646, 1129–1137 (2025). https://doi.org/10.1038/s41586-025-09581-z

News & media

Women portrayed as younger than men online, and AI amplifies the bias

October 8, 2025

In a sweeping study published today in Nature, researchers at UC Berkeley Haas, Stanford, and Oxford/Autonomy University documented extensive age and gender distortion across online media—and found that common algorithms are amplifying the bias.

UC Berkeley study finds Google Images and ChatGPT misrepresent women as younger than men in workforce

Oct 13, 2025

Despite no systemic age gap in the workforce, online databases and AI platforms depict women as younger than men across professions, according to a study conducted at UC Berkeley and published in Nature last week.

Study Overview

We present a deep learning-based approach to studying dynamic clinical behavioral regimes in diverse non-randomized healthcare settings. Our proposed methodology - deep causal behavioral policy learning (DC-BPL) - uses deep learning algorithms to learn the distribution of high-dimensional clinical action paths, and identifies the causal link between these action paths and patient outcomes. Specifically, our approach: (1) identifies the causal effects of provider assignment on clinical outcomes; (2) learns the distribution of clinical actions a given provider would take given evolving patient information; (3) and combines these steps to identify the optimal provider for a given patient type and emulate that provider's care decisions.

Study Results

Underlying this strategy, we train a large clinical behavioral model (LCBM) on electronic health records data using a transformer architecture, and demonstrate its ability to estimate clinical behavioral policies. We propose a novel interpretation of a behavioral policy learned using the LCBM: that it is an efficient encoding of complex, often implicit, knowledge used to treat a patient. This allows us to learn a space of policies that are critical to a wide range of healthcare applications, in which the vast majority of clinical knowledge is acquired tacitly through years of practice and only a tiny fraction of information relevant to patient care is written down (e.g. in textbooks, studies or standardized guidelines).

Working Paper: Knecht, J., Zink, A., Kolstad, J., & Petersen, M. (2025). Deep causal behavioral policy learning: Applications to healthcare. arXiv preprint arXiv:2503.03724. https://doi.org/10.48550/arXiv.2503.03724

News & media

UC Berkeley researchers, entrepreneurs create first healthcare AI model trained to understand clinical decisions

July 28, 2026

Experts in healthcare economics, computational health, biostatistics, and behavioral design from across the Haas School of Business and UC Berkeley came together to build a new kind of healthcare AI—trained on how doctors actually practice medicine.

Study Overview

How should the government respond to automation? We study this question in a heterogeneous agent model that takes worker displacement seriously. We recognize that displaced workers face two frictions in practice: reallocation is slow and borrowing is limited. We analyze a second best problem where the government can tax automation but lacks redistributive tools to fully alleviate borrowing frictions.

Study Results

The equilibrium is (constrained) inefficient and automation is excessive. Firms do not internalize that automation depresses the income of automated workers early on during the transition, precisely when they become borrowing constrained. The government finds it optimal to slow down automation on efficiency grounds, even when it does not value equity. Quantitatively, the optimal speed of automation is considerably lower than at the laissez-faire. The optimal policy improves efficiency and delivers meaningful welfare gains.

Working Paper: Martin Beraja and Nathan Zorzi, "Inefficient Automation," NBER Working Paper 30154 (2022), https://doi.org/10.3386/w30154.

Journal Publication: Martin Beraja, Nathan Zorzi, Inefficient Automation, The Review of Economic Studies, Volume 92, Issue 1, January 2025, Pages 69–96, https://doi.org/10.1093/restud/rdae019

Study Overview

Workers spend a significant amount of time learning how to make good decisions. Evaluating the efficacy of a given decision, however, can be complicated—for example, decision outcomes are often long-term and relate to the original decision in complex ways. Surprisingly, even though learning good decision-making strategies is difficult, the strategies can often be expressed in simple and concise forms. Focusing on sequential decision making, we design a novel machine learning algorithm that is capable of extracting “best practices” from trace data and conveying its insights to humans in the form of interpretable “tips.”

Study Results

Our algorithm selects the tip that best bridges the gap between the actions taken by human workers and those taken by the optimal policy in a way that accounts for which actions are consequential for achieving higher performance. We evaluate our approach through a series of randomized controlled experiments where participants manage a virtual kitchen. Our experiments show that the tips generated by our algorithm can significantly improve human performance relative to intuitive baselines. In addition, we discuss a number of empirical insights that can help inform the design of algorithms intended for human-AI interfaces. For instance, we find evidence that participants do not simply blindly follow our tips; instead, they combine them with their own experience to discover additional strategies for improving performance.

Journal Publication: Hamsa Bastani, Osbert Bastani, Wichinpong Park Sinchaisri (2025) Improving Human Sequential Decision Making with Reinforcement Learning. Management Science 72(1):733-755. https://doi.org/10.1287/mnsc.2022.02455

Study Overview

This project was formerly titled: Gender-Differentiated Digital Credit Algorithms Using Machine Learning 

Despite the promise of FinTech lending to expand credit access to populations without a formal credit history, FinTech lenders primarily lend to applicants with a formal credit history and rely on conventional credit bureau scores as an input to their algorithms.

Study Results

Using data from a large FinTech lender in Mexico, we show that alternative data from digital transactions through a delivery app are effective at predicting creditworthiness for borrowers with no credit history. Using account-by-month level data on revenues and costs, a machine learning model predicting profits generates similar profits as a model predicting default.

Intervention: AI model that differentiates creditworthiness between men and women

Intervention Partner: RappiCard

Populations: unbanked and underserved

Working Paper: Laura Chioda & Paul Gertler & Sean Higgins & Paolina C. Medina, 2024. "FinTech Lending to Borrowers with No Credit History," NBER Working Papers 33208, National Bureau of Economic Research, Inc.

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

News & media

There’s an easy way to make lending fairer for women. Trouble is, it’s illegal.

November 15, 2019

Preliminary results from an ongoing study funded by the UN Foundation and the World Bank are once again challenging the fairness of gender-blind credit lending. The study found that creating entirely separate creditworthiness models for men and women granted the majority of women more credit.

Gender-Differentiated Credit Scoring: A Potential Game-Changer for Women

February 27, 2020

The Alliance spoke to Sean about this research and the significant impact the model potentially could have on women’s ability to access credit.

Study Overview

Recent scholarship has suggested that artificial intelligence (AI) technology and autocratic regimes may be mutually reinforcing. We test for a mutually reinforcing relationship in the context of facial-recognition AI in China. To do so, we gather comprehensive data on AI firms and government procurement contracts, as well as on social unrest across China since the early 2010s.

Study Results

We first show that autocrats benefit from AI: local unrest leads to greater government procurement of facial-recognition AI as a new technology of political control, and increased AI procurement indeed suppresses subsequent unrest. We show that AI innovation benefits from autocrats’ suppression of unrest: the contracted AI firms innovate more both for the government and commercial markets and are more likely to export their products; noncontracted AI firms do not experience detectable negative spillovers. Taken together, these results suggest the possibility of sustained AI innovation under the Chinese regime: AI innovation entrenches the regime, and the regime’s investment in AI for political control stimulates further frontier innovation.

Working Paper: Martin Beraja, Andrew Kao, David Y. Yang, and Noam Yuchtman, "AI-tocracy," NBER Working Paper 29466 (2021), https://doi.org/10.3386/w29466.

Journal Publication: Martin Beraja, Andrew Kao, David Y Yang, Noam Yuchtman, AI-tocracy, The Quarterly Journal of Economics, Volume 138, Issue 3, August 2023, Pages 1349–1402, https://doi.org/10.1093/qje/qjad012

Study Overview

Overprecision is the excessive certainty in the accuracy of one’s judgment. This article proposes a new theory to explain it. The theory holds that overprecision in judgment results from neglect of all the ways in which one could be wrong. When there are many ways to be wrong, it can be difficult to consider them all. Overprecision is the result of being wrong and not knowing it. This explanation can account for why question formats have such a dramatic influence on the degree of overprecision people report. It also explains the ubiquity of overprecision not only among people but also among artificially intelligent agents.

Journal Publication: Moore, D. A. (2023). Overprecision is a property of thinking systems. Psychological Review, 130(5), 1339–1350. https://doi.org/10.1037/rev0000370

Study Overview

The data revolution in low- and middle-income countries is quickly transforming how companies approach emerging markets. As mobile phones and mobile money proliferate, they generate new streams of data that enable innovation in consumer finance, credit, and insurance. Already, this new generation of products are being used by hundreds of millions of consumers, often to use financial services for the first time. However, the collection, analysis, and use of these data, particularly from economically disadvantaged populations, raises serious privacy concerns. This white paper describes a research agenda to advance our understanding of the problem and solution space of data privacy in emerging market fintech and financial services.

Study Results

We highlight five priority areas for research: conducting comprehensive landscape analyses; understanding local definitions of "data privacy''; documenting key sources of risk, and potential technical solutions (such as differential privacy and homomorphic encryption); improving non-technical approaches to data privacy (such as policies and practices); and understanding the tradeoffs involved in deploying privacy-enhancing solutions. Taken together, we hope this research agenda will focus attention on the multi-faceted nature of privacy in emerging markets, and catalyze efforts to develop responsible and consumer-oriented approaches to data-intensive applications.

Intervention: Privacy-enhancing technologies, policies, and practices

Working Paper: Blumenstock, Joshua E.; Kohli, Nitin (2023): Big Data Privacy in Emerging Market Fintech and Financial Services: A Research Agenda. arXiv:2310.04970

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

Study Overview

We study the shifts in U.S. firms' workforce composition and organization associated with the use of AI technologies. To do so, we leverage a unique combination of worker resume and job postings datasets to measure firm-level AI investments and workforce composition variables, such as educational attainment, specialization, and hierarchy.

Study Results

We document that firms with higher initial shares of highly-educated workers and STEM workers invest more in AI. As firms invest in AI, they tend to transition to more educated workforces, with higher shares of workers with undergraduate and graduate degrees, and more specialization in STEM fields and IT skills. Furthermore, AI investments are associated with a flattening of the firms' hierarchical structure, with significant increases in the share of workers at the junior level and decreases in shares of workers in middle-management and senior roles. Overall, our results highlight that adoption of AI technologies is associated with significant reorganization of firms' workforces.

Working Paper: Tania Babina, Anastassia Fedyk, Alex X. He, and James Hodson, "Firm Investments in Artificial Intelligence Technologies and Changes in Workforce Composition," NBER Working Paper 31325 (2023), https://doi.org/10.3386/w31325.

Study Overview

Research on regulating emerging financial technologies ("FinTech") has been siloed to individual branches. Instead, this paper present a high-level view of various FinTech branches and analyzes the economic incentives of each. By focusing on the dynamics and parallels between the branches, the paper offers new insights for optimal regulation that balances the costs and benefits as use cases expand.

Study Results

Decentralized Finance (DeFi) which combines advances from the AI and blockchain branches, reduces the cost of coordinating complex financial services. Yet the efficiency gains intertwine with potential legal risks associated with liability, financial crime, dispute resolution, jurisdiction, and taxes. To ensure financial stability, effective regulatory solutions include adapted definitions and safe harbors, regulatory sandboxes, self-regulatory organizations, and/or policing misleading characterizations (e.g., regarding the extent of decentralization or agreed to data uses).

Intervention: Regulation

Working Paper: Grennan, Jillian, FinTech Regulation in the United States: Past, Present, and Future (August 31, 2022). Available at SSRN: https://ssrn.com/abstract=4045057

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)

Study Overview

In this review, we discuss changes resulting from both big data and non-traditional data in the financial industry. We explore ways these innovations are shaping markets and companies, but put greater focus on the consequences - both positive and negative - for end consumers. We aim to cast a wide net over offerings for both developed and emerging markets, as well as both products aimed at low-income, low-literacy individuals and those intended for more educated, wealthy, and financially savvy consumers.

Study Results

Analysis suggests there are many promises and potential risks - such as discrimination caused by bias in data used for credit scoring or consumers' behavior to game systems - related to data-driven fintech innovation. In many cases, better oversight and further research can help to address the concerns and ensure benefits are more widely distributed. As these technologies become more ubiquitous, the challenge will be for regulation and rigorous evaluation of impacts to keep pace with rising innovation.

Populations: General

Working Paper: Chioda, Laura and Gertler, Paul and Macdonald, Isabel and Wellsjo, Alexandra, Novel Use of Data in Fintech (June 02, 2022). Available at SSRN: https://ssrn.com/abstract=

IBSI Initiative or Center: Lab for Inclusive FinTech (LIFT)