Thomas H. Costello — Psychologist Carnegie Mellon University · Pittsburgh
I study the cognitive structures hidden beneath our stated issue positions — viewpoints — what they look like, what makes them change, and what follows from the fact that generative AI has begun to reshape the informational network they are built from.
Science · Cover story · Vol. 385
Evidence-based dialogues with an AI model reduced participants’ conspiracy beliefs by roughly 20 percent on average; the effects persisted for at least two months and generalized across a wide range of conspiracies, including strongly held and identity-relevant beliefs.
Read moreNature Medicine · 2025
Many warn, with reason, that AI will amplify misinformation or atrophy our critical faculties. In this essay I make the case for cautious optimism about what these systems could do for the information ecosystem.
Read morePublic tool
The public version of the chatbot from the Science study: anyone can describe a conspiracy theory they believe and talk it through with an AI that responds with evidence. It has now held more than 150,000 conversations with members of the public.
Try itI am an Assistant Professor in the Department of Social and Decision Sciences at Carnegie Mellon University, where I direct the Viewpoints Lab, with affiliations in the Human-Computer Interaction Institute, the Institute for Complex Social Dynamics, and the Center for Collaboration Science. My research uses tools from political psychology, decision science, human-computer interaction, and AI to understand how people form, structure, and change social and political beliefs.
I completed my PhD at Emory University under Scott Lilienfeld and Arber Tasimi, then a postdoctoral fellowship at MIT Sloan with David Rand and Gordon Pennycook, and spent a year on the faculty at American University before moving to Carnegie Mellon.
More about me
Familiar psychological categories — authoritarianism, cognitive and ideological rigidity, extremism — are often too coarse, polysemic, and multiply determined to explain political cognition. Two people can receive the same score on an authoritarianism scale, a conspiracy scale, or an attitude item for almost opposite reasons, and standard instruments hide exactly the heterogeneity that determines when and why minds change.
Large language models can serve as scalable semi-structured interviewers and experimental confederates: eliciting a belief in a person’s own words, then addressing their reasons with arguments targeted at the epistemic assumptions and evidence behind that specific claim — what I term epistemic personalization. In mechanism experiments, the evidence does the work; persuasion collapses when the model must argue without facts.
Persuasive capacity is fundamentally dual-use: in our experiments, models instilled false beliefs about as easily as they dispelled them. Meanwhile, chatbots have become a daily thinking infrastructure for hundreds of millions of people, where the more consequential scenario may be ambient exposure — answers whose defaults subtly push users toward particular conclusions over weeks or months.