Patrick Liu

A persistent puzzle in the study of public opinion is why political information often produces minimal attitude change despite reliably influencing beliefs. We argue that this duality reflects belief relevance—the extent to which specific beliefs bear on attitudes. Using conversations with large language models (LLMs), we elicit deeply held issue attitudes and the “focal beliefs” people use to justify those attitudes. We then randomly assign participants to receive an LLM-generated counterargument targeting either their focal belief, an issue-relevant but unmentioned belief (“distal belief”), or a placebo. In experiments with two large online convenience samples, counterarguments targeting the aforementioned beliefs successfully decrease belief strength, with effects persisting after 10 days. More importantly, focal belief counterarguments produce larger and more durable attitude change than distal counterarguments. These findings suggest that political information can successfully shift political attitudes and provide new evidence for the role of belief relevance in persuasion.

A long-standing debate in political psychology considers whether individuals update their beliefs and attitudes in the direction of evidence or grow more confident in their convictions when confronted with counter-attitudinal arguments. Though recent studies have shown that instances of the latter tendency, which scholars have termed attitude polarization and “belief backfire,” are rarely observed in settings involving hot-button issues or viral misinformation, we know surprisingly little about how participants respond to information targeting deeply held attitudes, a key condition for triggering attitude polarization. We develop a tailored experimental design that measures participants’ core issue positions and exposes them to personalized counter-attitudinal information using the large language model GPT-3. We find credible evidence of attitude polarization, but only when arguments are contentious and vitriolic. For lower valence counter-attitudinal arguments, attitude polarization is not detected. We conclude by discussing implications for the study of political cognition and the measurement of attitudes.

Theories in political science rest on latent constructs—political convictions, issue priorities, and identities—that vary meaningfully from person to person and resist a single operationalization. Faithful tests of these theories are difficult to achieve with standard experimental designs that present identical stimuli to all participants. This Element introduces tailored experiments as a design approach that more closely approximates the idiosyncratic ways individuals engage with their social and political environments. Advances in generative artificial intelligence, most notably large language models, have expanded the range of theoretical questions these designs can help answer and enabled implementation at scale. We situate tailored experiments within the potential outcomes framework and discuss requisite assumptions for causal inference. Through several case studies, we show how these designs enhance construct validity and yield new theoretical insights. We also examine technical challenges, ethical considerations, and other dimensions guiding their implementation. The Element concludes by considering how emerging capabilities and agentic models may further transform experimental political science.

Citizens increasingly report censoring their political beliefs to avoid backlash from co-partisans, yet the prevalence and origins of intra-party social pressure remain largely undocumented. This oversight obscures a critical question: is avoidance of political conversation driven by partisan norms that police dissent? Using a nationwide survey experiment during the 2024 U.S. primaries (N = 17,691), we find that partisans overestimate the likelihood of social sanctions for expressing their views. Experimentally reducing these exaggerated fears significantly reduces self-censorship among co-partisans for both loyalists and dissenters, affirming that intra-party disagreements are divisive enough to generate fears of social repercussions on all sides. Nonetheless, dissenters are substantially less willing than loyalists to discuss their preferences with co-partisans. Our findings suggest that this asymmetry may stem in part from partisan norms that raise the social costs of dissent but is also explained by minorities possessing fewer positive motivations to persuade intra-party opponents.

Political microtargeting has attracted widespread concern, particularly as large language models (LLMs) make it easier to generate personalized messages at scale. Yet existing studies find that ads tailored to voters’ demographic or personality characteristics seldom outperform a single pre-tested, best-performing message. In a two-wave pre-registered experiment, we compare the persuasive effects of LLM-generated audio campaign advertisements tailored to respondents’ demographics, Big Five personality traits, or self-identified issue priorities against an affordability advertisement identified as a top-performing message during the 2024 election. Consistent with prior work, demographic- and personality-based microtargeting using LLMs produce mixed effects on candidate choice in a hypothetical election setting. In stark contrast, ads tailored to respondents’ personally important issues increased candidate support by over 10 percentage points relative to a generic best-performing ad—equivalent to roughly one-fourth the effect of shared partisanship. These findings establish issue-based microtargeting as a meaningful upper bound for evaluating other targeting strategies, demonstrate that personalized, AI-generated campaign materials can be highly persuasive when grounded in voters’ actual priorities, and offer new evidence for the capacity of issue priorities to compete with partisanship in competitive contexts using a novel audio-based conjoint experimental design.

Why do anti-democratic candidates endure and sometimes thrive in American politics, despite broad public opposition to many anti-democratic behaviors? Existing explanations focus on voters’ failure to punish anti-democratic candidates because they agree with their policy stances. We propose and empirically evaluate a different mechanism, centered on voter learning: violations of democratic rules act as a signal of a candidate’s policy positions and/or strong resolve. We propose a formal model and evaluate its implications using survey experiments that manipulate both the presence of democratic violations—such as election denial and threats of violence against opposing partisans—and the amount of information voters receive about candidates’ policy stances and partisan loyalty. Our results show that voters’ propensity to defect from anti-democratic candidates is less responsive to their ideology when they have more information. In line with our proposed mechanism, these effects are driven by differential inferences about future behavior once in office.