Large Language Models Polarize Ideologically but Moderate Affectively in Online Political Discourse
Gavin Wang Srinaath Anbudurai # Oliver Sun # Xitong Li Lynn Wu # Equal contribution. · 2026
Read Large Language Models Polarize Ideologically but Moderate Affectively in Online Political Discourse by Gavin Wang Srinaath Anbudurai # Oliver Sun # Xitong Li Lynn Wu # Equal contribution. (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-20238.txt.
Introduction
Gavin Wang and four colleagues study how the release of ChatGPT changed political talk on Reddit's largest political forum. They examine ideological and emotional patterns across millions of comments. Their analysis shows that ideological polarization grew after the release: liberal-leaning authors posted still more liberal comments, and conservative-leaning authors posted still more conservative ones.
The title states the paired finding: models polarize users ideologically but moderate them affectively, in tone. That split matters for how platforms judge AI's effect on public debate. A tool can calm the mood of a thread while quietly pushing its sides further apart.
Catalog record
| CREATOR | Gavin Wang Srinaath Anbudurai # Oliver Sun # Xitong Li Lynn Wu # Equal contribution. |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 128 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-20238
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Large Language Models Polarize Ideologically but Moderate Affectively in Online Political Discourse Gavin Wang Srinaath Anbudurai # Oliver Sun # Xitong Li Lynn Wu # Equal contribution. Abstract The emergence of large language models (LLMs) is reshaping how people engage in political discourse online. We examine how the release of ChatGPT altered ideological and emotional patterns in Reddit’s largest political forum. Analysis of millions of comments shows that ChatGPT intensified ideological polarization: liberal-leaning authors posted increasingly liberal comments, while conservative-leaning authors posted increasingly conservative comments. Multiple falsification tests suggest that these findings are unlikely to be driven by contemporaneous events, such as the 2022 U.S. midterm elections, or by broader platform-wide trends in political polarization. Mechanism tests show that this shift does not stem from the creation of more persuasive or ideologically extreme original content using LLM. Instead, it originates from the tendency of LLM-assisted comments to echo and reinforce the original post’s viewpoint, a pattern consistent with algorithmic sycophancy. Yet, despite growing ideological divides, affective polarization, measured by hostility and toxicity, declined. These findings reveal that LLMs can simultaneously deepen ideological separation and foster more civil exchanges, challenging the long-standing assumption in literature that extremity and incivility necessarily move together. Keywords: Large Language Models, ChatGPT, Political Polarization, Liberal, Conservative, Affective Polarization 1 Introduction Incivility has long been understood to co-evolve with ideological polarization. A substantial body of work argues that rising ideological polarization is systematically accompanied by rising incivility ( Chen and Rohla, 2018 ; Santoro and Broockman, 2022 ; Zimmerman et al., 2022 ; Boxell et al., 2024 ) . While this literature establishes a strong association in human-authored discourse, its premises become uncertain in the AI era, where much of the information environment is increasingly mediated by large language models (LLMs). This shift motivates a re-examination of the ideological polarization-incivility nexus in the context of AI. This study examines how the release of ChatGPT impacted political discourse on social media platforms. The advent of LLMs, particularly OpenAI’s ChatGPT, has transformed how individuals access information and create content. These models, developed through extensive training on diverse internet content, can generate human-like text, answer queries, and assist in a variety of tasks ( Noy and Zhang, 2023 ) . While their applications span domains such as education, healthcare, and customer service ( Bastani et al., 2025 ; Brynjolfsson et al., 2025 ; Zöller et al., 2025 ) , their influence on political discourse remains comparatively underexplored. This omission is critically important because LLMs are not neutral conduits of information: prior studies suggest that ChatGPT exhibits a “pro-environmental, left-libertarian orientation” ( Rozado, 2023 ; Rutinowski et al., 2024 ) , and, in the U.S. context, a tendency to favor Democratic positions ( Motoki et al., 2024 ) . Such biases may shape not only the content users generate using these tools, but also the ideological direction and tone of politically charged online discussions. These concerns are particularly consequential on social media platforms, which have become central arenas for political discussion. Social media not only reflects public opinion but also actively shapes civic engagement and patterns of political polarization ( Waller and Anderson, 2021 ; Wojcieszak et al., 2022 ; Mosleh et al., 2024 ) . These platforms enable rapid information dissemination, foster echo chambers, and facilitate interactions among users with varying ideological leanings ( Bail et al., 2018 ; Flamino et al., 2023 ; Mamakos and Finkel, 2023 ; Nyhan et al., 2023 ) . As political conversations increasingly migrate online, the nature and quality of discourse in these spaces can affect democratic legitimacy, shape policy debates, and even influence electoral outcomes ( Grinberg et al., 2019 ; Siev and Petty, 2024 ) . By lowering the cost of producing political content and shaping the form of online interactions, LLMs may alter the ideological composition and tone of political discourse on social media. Understanding how LLMs affect political discourse on social media is thus crucial for assessing their broader implications for democratic governance and societal cohesion. We focus on the political discourse on Reddit which is one of the largest and most ideologically diverse social media platforms, and has been broadly examined in prior research on political polarization ( Waller and Anderson, 2021 ; Mamakos and Finkel, 2023 ) . We assemble large-scale, detailed comment-level data from Reddit’s most active political forum, including detailed post and comment content, timestamps, and author information. Leveraging this data, we assess changes in Reddit users’ political expressions before and after the introduction of ChatGPT. Following the practice of previous literature ( Gentzkow and Shapiro, 2010 ; Greenstein and Zhu, 2012 ; Greenstein and Zhu, 2016 ) , we measure the political slant of each comment on a continuous scale by mapping the frequency of words and phrases to their ideological usage in congressional speeches, and calculate each author’s political stance at different times. As a contribution, rather than relying on the 2005 record as in prior work, we construct our dictionary from the 2021 Congress Record to more accurately capture contemporary political language. We also employ several state-of-the-art indicators that have been widely used in prior research to assess the likelihood that a Reddit comment was generated with the assistance of LLMs, including comment length, passive verbs, perplexity score and human score ( Radford and Wu, 2019 ; Solaiman et al., 2019 ; Simchon et al., 2023 ; Muñoz-Ortiz et al., 2024 ) . Because direct information on whether a particular Reddit comment was generated using LLMs is unavailable, our analysis relies on established linguistic and model-based indicators that estimate the probability of LLM involvement in content creation rather than definitively identifying AI-generated content. Although these methods cannot determine with complete certainty whether any individual comment was generated with LLM assistance, they have been shown to capture meaningful variation in LLM involvement and are well-suited for large-scale population-level inference ( Shan and Qiu, 2026 ) . In a corpus comprising millions of comments, idiosyncratic classification errors are unlikely to systematically generate the aggregate patterns we document. Our empirical results indicate that the release of ChatGPT significantly intensified political polarization on the platform: comments posted by previously liberal-leaning users shifted further toward the liberal end of the ideological spectrum, while comments posted by previously conservative-leaning users shifted further toward the conservative end. The magnitude was substantively meaningful: the impact of ChatGPT’s release on average was equivalent to shifting a median liberal (conservative) author into the top 30% most extreme segment of their respective ideological spectra. We also investigate how the release of ChatGPT affected authors near the ideological center and those at the margins of the liberal–conservative boundary. We find a systematic leftward shift across these groups, though not substantial in magnitude. Multiple falsification tests suggest that our findings are unlikely to be driven by contemporaneous events, such as the 2022 U.S. midterm elections, or by broader platform-wide trends in political polarization. Despite this amplification of ideological alignment within discussions, we observe a clear decline in affective polarization ( Iyengar et al., 2012 ; Boxell et al., 2024 ) —the extent to which individuals express hostility or animosity toward opposing political groups. Measures of civility on the platform, captured through comment toxicity and hostility, improved