In this article, we speak with Prof. Aditya Dasgupta, a political science faculty who integrates AI into social science research. Adi is the director of the Political Economy of Agriculture and Rural Societies (PEARS) Lab, where he explores how machine learning, neural networks, and computer vision can transform data collection and analysis. From tracking farmer protests in India using AI-driven text classification to predicting voter turnout based on satellite imagery in Mexico, his work showcases the power of AI in political research.
AI and computer vision: neural networks for processing and analyzing data
“Our political science department takes a quantitative approach,” Adi says, “requiring well-measured variables for statistical analysis.” Traditionally, this meant manually collecting data—reading documents, coding images, and analyzing archives. At the PEARS Lab, Adi has automated much of this process using custom-trained algorithms. Whether analyzing thousands of newspaper articles or coding satellite images, AI efficiently transforms unstructured text and imagery into structured datasets, making large-scale analysis far more feasible.
From the Green Revolution in India to predicting voter turnout in Mexico
One fascinating project Adi describes is examining the political impact of the Green Revolution in India. The Green Revolution introduced new agricultural technologies in the 1960s, which helped India achieve self-sufficiency in rice and wheat production. However, it also transformed the country’s political landscape. “It changed the nature of agriculture from subsistence farming to modern industrial farming, using extensive irrigation, pesticides, and fertilizers. As a result, farmers began engaging much more politically because all those resources were subsidized or provided by the state,” he explains.
To study this political engagement, his team set out to measure collective action by farmers—tracking protests, rallies, and other demonstrations over time. They analyzed thousands of archived articles from The Times of India, using keyword searches to identify relevant reports dating back to the 1960s. “We had over 10,000 PDFs, but only a small fraction actually described the events we were interested in.” To handle this volume, they took a two-pronged AI-driven approach
“One method was simple—we used ChatGPT. We wrote a script that transcribed article text and sent it to the ChatGPT API to determine whether it contained information about farmer protests. If so, it extracted key details like location, date, and participants.” The second approach involved training a custom neural network. “Instead of using an off-the-shelf model, we had RAs manually code a sample of articles, then trained an algorithm to predict whether an article was useful based on its word sequences.
This AI-assisted workflow allowed the team to process an otherwise overwhelming dataset in a single semester. “A massive undertaking, made much less massive because of new tools.”
Adi further describes how he is increasingly exploring AI for inference, aiming to identify causal relationships in political science research. “One project I can talk about is a paper with my colleague, Prof. Tesalia Rizzo Reyes, who specializes in Mexican politics,” he shares. Together, they investigated whether imagery of polling places could predict voter turnout
Using Python, they retrieved latitude and longitude coordinates of polling stations and pulled both front-facing and overhead satellite images via the Google Maps API. “We trained a convolutional neural network to predict voter turnout just from images of buildings,” he explains. Testing the model on unseen data, they found that polling place imagery alone could explain about 30% of the variation in voter turnout. “That told us we were onto something—these images contain a lot of information about whether people actually vote.”
The next step was understanding why. Using explainable AI techniques, they analyzed which visual features the model relied on. “We found that larger, more spacious buildings were associated with higher turnout, while cramped spaces correlated with lower turnout. Polling places with tree cover and green space saw higher participation, while those in dense, concrete-heavy environments had lower turnout. Government buildings—such as schools—were linked to higher turnout compared to private homes or shops.”
These insights made intuitive sense. “In the context of a pandemic, people likely prefer voting in a big space rather than a cramped one. In extreme heat, shade and green space might make a difference. And in areas with security concerns, voters might feel safer casting their ballots in an official building rather than a private shop.” This project demonstrates how an open-ended, data-driven approach using AI can uncover meaningful political patterns that align with real-world behavior.
Bridging Political Science and AI: A Distinct Advantage in Graduate Education
To go beyond merely using tools like ChatGPT, Adi took a unique approach to self-education. “When I was in graduate school, there were no courses on these tools,” he explains. During a pre-tenure sabbatical at Stanford’s Hoover Institution in 2020-2021, he faced a choice: focus on publishing more articles or invest time in learning AI. “I took the gamble to really learn about neural networks, computer vision, and these emerging technologies, hoping it would keep me up to date in the long run.” This decision has since shaped his research and teaching, equipping him with skills that few in his field possess
And he continues to pay this forward by actively incorporating AI into graduate education. Last spring, he taught the university’s first graduate seminar on AI for Social Science, introducing students to neural networks and deep learning. The PEARS lab is equipped with a powerful deep learning computer, and he encourages students to take engineering courses, particularly in computer vision. “These are things I wasn’t doing in grad school—this is a whole new frontier for the next generation.”
The seminar assumed some familiarity with coding and statistics but started from the basics, gradually building toward complex models. It was hands-on, and many students are now using AI in their dissertations.
“Writing is Thinking”
Adi is teaching an undergraduate course this spring titled The Politics of Poverty and Prosperity, which explores global economic development from the 17th century to the present. “We look at why some countries became rich while others remained relatively poor and examine the role of political institutions in shaping these different trajectories,” he explains. The course has roots in his own academic journey—he first encountered these ideas while serving as a TA for a similar class at Harvard, taught by James Robinson, who recently won the Nobel Prize.
In this context, he is thinking critically about AI’s impact in undergraduate writing. “Writing is thinking,” he says. “It would be terrible if students stopped writing their own essays and relied too heavily on large language models. You're not just outsourcing an assignment—you’re outsourcing the thinking.” To address this, his undergraduate course will require deep engagement with primary sources, making automation more difficult. “ChatGPT struggles with citations—it makes up sources. By requiring detailed references to class readings, we create a safeguard against students relying entirely on AI.
As AI continues to transform research, governance, and daily life, Adi remains committed to equipping students with the critical skills and knowledge they need to navigate—and shape—this rapidly evolving landscape. “Universities need to prepare students—both undergraduates and graduates—for this new world,” he explains.
Aditya’s work shows that AI isn’t just changing research—it’s changing the way we see the world. By blending technology with political science, he reminds us that staying curious, adapting to new tools, and questioning the systems around us are key to making sense of the future.
Interested in contributing? Email: sghosh@ucmerced.edu
