Thomas R. Davidson

Why have populist and far-right actors been so successful at building large online audiences? To what extent is online activism driven by offline events — protests, elections, terrorist attacks — versus endogenous processes within social media itself? Do extremists benefit from platform affordances, particularly ranking and recommendation algorithms? I bring together theories from political sociology, social movements, and public-opinion scholarship with computational methods applied to social-media, news, and other text data.

This work includes case studies focused on the United Kingdom and comparative studies of political parties across Europe. In a recent article in Mobilization, I argue that algorithmic feedback loops enable social-movement actors to generate and sustain attention. In Political Communication, I show that European populist parties attract more Facebook engagement than other parties and that this advantage is growing. A recent article in the Journal of Ethnic and Migration Studies analyzes how parties adapted their language during the Syrian refugee crisis, and a forthcoming article in Comparative Political Studies examines the COVID-19 surge in support for right-wing populists in Europe.

Ongoing projects include the use of generative AI to create visual propaganda and agent-based models of how algorithmic ranking shapes online activism and support for extremists.

A second area of my research focuses on identifying and understanding hate speech on social media. My early work drew the distinction between hate speech and other forms of offensive language, showing how conflating the two produced false positives in machine-learning classifiers. This research was covered in Wired, Tech Republic, and New Scientist.

I subsequently developed theoretical work on the dimensions of hateful and abusive language (paper) and examined racial bias in detection systems — demonstrating that classifiers are more likely to label tweets written in African-American English as hateful than similar tweets in Standard American English (paper). This work was covered in Vox. A chapter in the Oxford Handbook on the Sociology of Machine Learning provides an overview of this research and identifies directions for further inquiry.

I have conducted experimental research on how social context influences judgments about hateful or abusive content, supported by a Foundational Integrity Research award from Meta. A paper in Nature Human Behaviour compares parallel experiments with human subjects and vision language models, showing how AI can make context-sensitive moderation decisions that align with human judgment. A related paper in ACL Findings 2023 combines LLMs and experiments to understand how social contexts shape perceptions of toxicity. I have spoken about this research at policy dialogues organized by the Organization of American States and a content-moderation working group at the European Commission.

I also study how computational methods can be applied more generally in sociological research. Most recently I have published several articles examining the uses of large language models and generative AI for sociology. My 2024 article in Socius showcases the methodological possibilities created by generative AI. In a 2025 article in Sociological Methods & Research, my advisee Youngjin Chae and I evaluate LLMs for text classification and offer recommendations for best practices. With Daniel Karell at Yale, I organized the first workshop on Generative AI and Sociology in 2024 and guest-edited a special issue of SMR on the subject.

In earlier work, I explored machine learning to study social processes. As part of the Fragile Families Challenge, I examined whether neural networks accurately predict social outcomes and whether these black-box models can be amenable to sociological explanation. The methods do not substantially outperform linear regression, but they enable inductive identification of important variables. The analysis appeared in Socius, and the mass-collaboration results were published in PNAS.

Beyond academia, I have applied these methods in industry settings. In 2016 I was an Eric and Wendy Schmidt Data Science for Social Good Fellow at the University of Chicago, where I helped develop an early-warning system for police misconduct (project; coverage in the Chicago Tribune, NPR, Mother Jones, the Economist, and Forbes). In 2017 I worked at Civis Analytics in Chicago, building an NLP tool to monitor political discussions on Twitter. In 2018 I was a Core Data Science intern at Facebook, studying misinformation sharing and helping evaluate tools for detecting coordinated inauthentic behavior.