Dissertation Title: Language Models to Understand, Reflect, and Address Social Fragmentation

Abstract:
Social fragmentation in political and civic life has steadily worsened over the past several decades. While technologies like social media initially promised to foster connection, they have instead often deepened divisions and amplified polarization. With the rise of generative artificial intelligence, there is an opportunity to shape a powerful general-purpose technology to counteract this fragmentation—but only if we deliberately design it to do so. This thesis explores how language models can be used to understand social fragmentation, how these models themselves can reflect the biases and divisions in political life, and finally how, if designed intentionally, they can be used to strengthen our social fabric.

In the first part of this thesis, I investigate how language models can be used to analyze political polarization at scale on social media. I find a consistent rise in the use of polarizing discourse, especially among political elites, suggesting that these platforms have grown increasingly ineffective at fostering constructive political dialogue.

In the second part of this thesis, I turn to the language models themselves, examining how they consistently exhibit a left-leaning political bias, even when trained solely on data optimized for truthfulness. This tendency is also evident when models are tasked with representing short-term and long-term preferences: those designed to act in an individual’s long-term interest align more closely with expert consensus but display stronger left-leaning bias on subjective or value-laden issues. These findings reveal several trade-offs in using language models to represent human preferences and underscore the risks of delegating political or moral judgment to “digital representatives” that may further distance individuals from the democratic process and from one another.

Building on these insights, the final part of this thesis proposes an alternative approach to the civic use of language models. Rather than substituting genuine participation with “digital clones” designed to replicate individual preferences, I show how language models can be employed to elicit authentic beliefs and render them navigable in ways that foster mutual understanding. I introduce a scalable system for collective decision-making grounded in people’s voices and experiences that enhances process legitimacy and social cohesion. Finally, I adapt this system into a tool that cultivates the skill of consensus building through dynamic opinion predictions on salient policy issues.

While there are many concerns about the impact of generative AI on an already fragmented society, this thesis shows that, rather than replacing participation and further weakening social ties, such technology can be used to elicit and interpret authentic human voices and experiences, ultimately helping to strengthen our social fabric.

Committee members:
Deb Roy, Professor of Media Arts and Sciences, Director of MIT Center for Constructive Communication
Diyi Yang, Assistant Professor of Computer Science, Stanford University
Michiel Bakker, Assistant Professor of Contemporary Technology, MIT Sloan School of Management and Senior Research Scientist at Google DeepMind