Imagine a social-science study that, instead of recruiting human respondents, prompts an artificial intelligence system to generate the responses for them. That scenario is no longer hypothetical: large language models can now produce fluent, plausible, and often statistically passable simulations of human opinions. Eun Cheol Choi, a Ph.D. candidate in Communication at the University of Southern California, has built his research around a deceptively simple but urgent question: What actually survives, and what quietly disappears, when a human research participant is replaced by a simulated proxy, or what he calls a “silicon sample”? In an upcoming talk titled “Proxies for People: Social Networks, Misinformation, and Simulated Respondents,” Choi will present findings that caution against both uncritical adoption and outright dismissal of AI as a stand-in for human subjects. Rather than asking whether AI proxies are simply beneficial or detrimental, his work probes when they are trustworthy, what they fail to capture, and what standards researchers should demand before treating model-generated responses as evidence. The question matters beyond academic method. Across disciplines, researchers are beginning to use large language models as proxies for human judgments in surveys, experiments, and policy simulations. If those proxies distort the relationships that social science is designed to understand, then the resulting findings may look rigorous while being quietly false. Choi, a computational social scientist trained across communication and computing, uses his own domain—misinformation and social networks—as a critical testing ground to explore these themes, and his presentation promises to be as much a warning as a roadmap for the responsible use of generative AI in social research.
Choi’s research is organized around two connected threads. The first thread focuses on how the structure of people’s social networks shapes their belief in and sharing of misinformation. Conventional explanations of misinformation often emphasize individual psychology: cognitive biases, stubborn partisanship, or knowledge gaps. Choi’s work adds a crucial structural dimension. According to the abstract of his presentation, initial findings from a survey of U.S. adults show that susceptibility to misinformation is influenced by both individual attitudes and the structure of social networks. That means people are not isolated decision-makers; they are embedded in webs of relationships that determine which claims they see, which sources they trust, and whether they have contacts willing and able to correct them. A person with relatively strong critical thinking skills might still share a false claim because their network rewards sharing or because the same misinformation surrounds them from multiple directions. Another person, more vulnerable in terms of personal bias, might be protected by a heterogeneous network containing credible responders. Understanding such dynamics is essential to designing effective interventions. If misinformation spreads because it is reinforced within tightly connected clusters, then fact-checking alone may be insufficient; interventions may need to target specific network positions, bridge nodes, or information pathways. If susceptibility is partly a function of social context, then studies that treat individuals in isolation are missing a key part of the causal story. Choi’s work thus establishes a foundation: before researchers can decide whether AI can simulate human respondents, they first need to understand exactly how human respondents are shaped by the social structures surrounding them.
The second thread of Choi’s research turns to AI itself. Even as social scientists come to appreciate the importance of social networks, many are using large language models to simulate human participants, perhaps in hopes of reducing cost, increasing scale, or avoiding ethical complications. Choi has tested this practice directly. In his experiments, he takes actual survey respondents and prompts large language models to respond as those respondents might. He then compares the model-generated responses to the human responses, not only in terms of average opinions but also in terms of underlying patterns and relationships. His findings are sobering. Current large language models, he reports, tend to overemphasize individual-level tendencies and have difficulty replicating the relational structures that are critical to human behavior. In practice, a model might be quite good at reproducing a general distribution of attitudes on a topic. It might even match demographic differences in surface ways. But it tends to struggle with the things that matter most for social-scientific understanding: who knows whom, who trusts whom, how information moves across a network, and how peer influence shapes individual decisions. A simulation can therefore appear accurate on the surface while missing the very mechanisms that make social life social. The danger is especially pronounced in misinformation research. Belief in false claims is not simply a personal trait; it is dynamically related to communication contexts, social pressures, and the information environment. If AI-simulated respondents cannot reproduce those relational structures, studies that rely on them may produce distorted conclusions about why misinformation spreads, whose beliefs are vulnerable, and what kinds of correction might work. Choi’s findings suggest that using AI as a proxy without careful validation is not just an abstract methodological problem; it is a substantive threat to the validity of social research.
Against this backdrop, Choi advocates a more rigorous evaluation standard for AI-simulated respondents. He proposes that researchers assess AI responses across multiple dimensions of fidelity, following the criteria social scientists already use to disentangle complex relationships among cognition, behavior, and social context. It is not enough for a model to generate the right mean opinion on a Likert scale. A trustworthy simulation should also reproduce the relationships among variables, the ways attitudes depend on social position, the conditional patterns of behavior, and, ideally, the mechanisms that connect one factor to another. Choi calls for evaluation that is multidimensional: researchers should ask whether a simulated respondent captures not only what people think but how they think, how they interact with others, and how their beliefs would change under different network conditions. His presentation also addresses ongoing efforts to improve simulation accuracy. Large language models are improving rapidly, and better models, more careful prompts, and more informed validation protocols may eventually make AI proxies more faithful than they are today. Yet there are also fairness challenges that do not disappear simply because a model is more powerful. Because models are trained on data that represent some populations far better than others, they may simulate certain demographic or cultural groups more faithfully while flattening, caricaturing, or erasing others. If social scientists rely on these tools uncritically, they risk building a literature informed primarily by the voices of those who are easiest to simulate, while marginalizing precisely those communities that are underrepresented in training data. For Choi, fairness is therefore not an afterthought; it is central to any responsible decision to use generative AI as a substitute for human participants. He offers both a cautionary perspective on the limits of AI proxies and a constructive path forward: a set of standards that would help researchers decide not just whether a simulation is plausible, but whether it is valid for the specific question being asked.
Choi’s own background gives these questions unusual depth. He is a Ph.D. candidate in Communication at the University of Southern California, where he also earned an M.S. in computer science. This dual training is central to his identity as a computational social scientist, able to move between the formal languages of machine learning and the theoretical frameworks of communication research. Before coming to USC, he earned his M.A. and B.A. in communication from Seoul National University, and he worked at the SNU FactCheck Center, then South Korea’s largest fact-checking platform. That experience gave him firsthand knowledge of how misinformation manifests in real-world information environments, how fact-checkers labor to correct false claims, and why corrections sometimes fail to reach the people who need them most. Since 2022, Choi has served as a research assistant at the USC Information Sciences Institute, contributing to projects supported by the Defense Advanced Research Projects Agency and the National Science Foundation. He has also worked as a teaching assistant for courses in Data Science for Communication and Social Networks at USC. In addition to his academic research, he has built tools that help fact-checkers identify recurring misinformation using large language models, demonstrating his commitment to putting AI research into practical service. His work has been published in the journal Social Networks and presented at leading technical venues, including the International Conference on Machine Learning, the International AAAI Conference on Web and Social Media, and the ACM Web Conference. This combination of academic rigor, practical fact-checking experience, and technical fluency gives Choi a distinctive vantage point: he is neither naive about the promise of AI nor dismissive of its real dangers.
For the RIT community, Choi’s talk will be an opportunity to engage with these issues directly. He has said he looks forward to meeting RIT’s faculty and community and exchanging ideas about misinformation, social networks, and the responsible use of AI in social research. He welcomes conversation with colleagues in communication, computing, and cognitive science, and the presentation has been planned with accessibility in mind: ASL-English interpreters have been requested, and light refreshments will be provided. The talk arrives at a time when generative AI is being rapidly integrated into research workflows, sometimes with enthusiasm but often without a clear understanding of the limits of AI-generated data. Choi does not argue that AI should never be used to simulate human respondents. Instead, his message is one of accountability: before the social sciences scale up the use of “silicon samples,” researchers must ask, dimension by dimension, whether those simulated proxies are faithful enough for the specific research question at hand. If a model cannot reproduce the relational structure of human communities, then it cannot be trusted to answer questions about how misinformation travels through those communities. If it represents some populations better than others, then its outputs cannot be treated as universal. By proposing a multidimensional evaluation standard, Choi offers both a cautionary perspective and a constructive path forward. His work has significant implications for the responsible use of generative AI across the social sciences, and it invites a broader conversation about what we lose, and what we must demand, when we ask machines to serve as proxies for people.
