Featuring Amreeta Das, PhD student in Political Science at UC Merced.
From the Campaign Trail to the Research Lab
Before she began her PhD at UC Merced, Amreeta worked for a politician in India during a state election. The experience was exhilarating and chaotic—fast-paced, image-driven, and filled with behind-the-scenes strategizing. "I got to see firsthand how social media campaigns were built,” she recalls. “There’s a lot of thought behind every graphic you see online—the color, the framing, even who appears in the image. It’s never random."
That year of real-world campaigning shaped her dissertation topic. “I realized we didn’t have good scholarly tools to study this kind of visual communication,” she says. “People were analyzing speeches and text posts, but the real action was happening in images and videos.”
AI as a Research Lens
Amreeta’s research uses deep learning, computer vision, and generative AI to study how political messages are crafted and perceived. “Social media is no longer just about words,” she explains. “It’s visual, emotional, and fast. My question was: how do we study that systematically?”
Her approach draws from computer science methods—training neural networks to detect patterns across thousands of campaign visuals. “For me, the challenge was figuring out how AI can help political scientists analyze visual data at scale,” she says. “We’ve relied on surveys and text analysis for years, but AI opens a new window into how persuasion actually looks.”
From Pixels to Persuasion
Her first project focused on political campaign imagery in India. Using thousands of official party posts, Amrita trained computer-vision models to see if they could distinguish visuals from different parties—without any text input. “It worked,” she says, smiling. “The models reached high accuracy just from the images, which means there are consistent visual cues across parties—even when humans can’t easily articulate them.”
To find out what the AI was actually “seeing,” she used Explainable AI techniques that highlight which regions of an image drive predictions. “The models often focused on the party symbol or the leader’s face, which made sense,” she explains. “But they also picked up more subtle things—like background colors, crowd types, or where the national flag appeared, or didn’t.”
AI + Humans = Smarter Research
Amreeta emphasizes that even the most advanced models still need human oversight. “AI is powerful, but it makes mistakes,” she says. “You have to treat it like a very fast, sometimes overconfident intern.”
In her lab, she manages undergraduate research assistants who verify AI outputs and help clean and annotate data. “A lot of social-science data is messy,” she explains. “We have to digitize election records, fix errors, and match datasets before we can even start analysis.” Her students help build validation pipelines that combine automated scripts with manual checks—a workflow that mirrors how professional data scientists operate.
“The human-in-the-loop model is crucial,” she says. “AI can scale what we do, but it can’t replace reasoning.”
The Next Frontier
Now, Amreeta wants to look beyond official campaign messaging to what she calls the bottom-up campaigns—the videos, memes, and voice notes ordinary people share on encrypted platforms like WhatsApp. “That’s where virality really happens,” she says. “Messages evolve as people pass them along. Authenticity, emotion, and misinformation all blur together.”
Studying that kind of data is tricky, both ethically and technically. “Platforms like WhatsApp are encrypted, so collecting data responsibly takes collaboration with other researchers and careful design,” she explains. But she believes it’s essential. “If we want to understand how citizens shape political narratives, we can’t only study the top-down messages.”
Seeing the Invisible in Politics
Amreeta's work sits at the intersection of political science and computer vision—a rare bridge across disciplines that rarely speak to each other. “For me, AI is not about replacing analysis,” she says. “It’s about revealing what’s invisible to the human eye—patterns of persuasion, emotion, and power that we might otherwise miss.”
Her favorite tools reflect that diversity: Claude for coding and structured reasoning, ChatGPT for writing and editing, Mistral for image tasks, and Google’s Deep Research for background synthesis. “Each one is like a different lab instrument,” she says. “You just need to know what to use, and when.”
In the end, Amreeta believes that understanding politics in the digital age means learning to read both the words and the images. “The stories we see are data, too,” she says. “AI just helps us read them more clearly.”
