Aaron Russell Kaufman

Language models are becoming an ordinary way to ask a political question, and there is widespread concern that their answers lean left. Nearly all evidence for this concern is national: each audit scores a model against a single country's survey respondents, most often American ones, and the resulting score confounds where the model stands with where the sampled population stands. Here we remove this confound by scoring models on item parameters estimated from 1.14 million human survey respondents in 141 countries and then held fixed, placing models, respondents and countries on one metric anchored to people rather than to any national sample. We pose 1,153 harmonized survey items to 65 models from 27 developers, at temperature zero and with no persona, and score the answers on four latent dimensions: social liberalism, economic left-right, democratic commitment and anti-elite sentiment. The quantity this measures is expressed ideology, the positions a model states when asked, rather than any claim about what these systems hold. We find, first, that all 65 models sit to the liberal side of the world human median on social liberalism, with the median model at the 31st percentile, and that the lean runs weaker in the same direction on the other three dimensions. We find, second, that the models resemble one another far more than people do: across three orders of magnitude in parameter count, and including open-weight, proprietary, reasoning-tuned and deliberately uncensored variants, the standard deviation of model positions is 11 to 33 percent of the human standard deviation on every dimension, and differently post-trained versions of the same base model stay within that band. Two further results caution against reading any single number too closely: changing which survey program supplies the items moves a model's measured position by 0.28 to 0.62 human standard deviations, two to nearly four times what the same change does to a country and more than the entire model-human gap on every dimension, and with the number of answered items held equal on both sides, the difference between models and human respondents in how consistently they place themselves on one ideological side shrinks to a modest residual. The direction of the liberal lean is robust across instruments; its magnitude is a property of the model and the questionnaire together.

Societies worldwide are described as increasingly polarized, yet issue polarization-how divided a society is over political positions-has been measured almost entirely on a single left-right axis developed in Western democracies. Whether that axis describes political division elsewhere is largely untested, because no survey measures political attitudes on multiple dimensions across many countries. Here, we harmonize 1,299 items from nine cross-national survey programs and 23 national election studies to place 1.14 million people from 141 countries on four dimensions: social liberalism, economic left-right, democratic commitment, and anti-elite sentiment. A single left-right axis does not describe political division in most of the world. The median within-country correlation between a person's social and economic positions is-0.015. Societies differ instead in what they are divided about, and they do so by region. European countries divide most often over social liberalism, East Asian and Arab countries over democratic commitment, Latin American countries over economic left-right, and African countries over anti-elite sentiment. Of the 81 countries measured on all four dimensions, 39 are most divided on a dimension no left-right scale would capture. Most political disagreement lies within countries rather than between them, at 58 to 88 percent of the variance. Demographic characteristics are only weakly associated with political attitudes, in directions that vary across countries. Ideological constraint, long treated as a property of mass publics, holds in nineteen of the eighty-three we can test, most strongly in the United States at +0.70, and there it is concentrated among the politically attentive. Understanding political conflict therefore requires asking not how polarized a society is but what it is polarized about. Our released item bank and fitted parameters place respondents from any future survey on the same scale without re-estimation, once that survey's items are harmonized to the bank.

Partisan gerrymandering is a familiar practice. But intra-partisan gerrymandering—one faction within a party designing districts to handicap another faction within the same party—has barely been noticed by courts or scholars. In jurisdictions dominated by a single party, though, intra-partisan gerrymandering is more impactful than its partisan counterpart. In these places, there’s no doubt which party will govern. What’s uncertain is which actors within this party will prevail. Intra-partisan gerrymandering matters precisely because it shapes the identity of the ruling intra-party coalition. In this Article, we first conceptualize intra-partisan gerrymandering. We define the practice, compare it to other forms of abusive redistricting, and explain how it unsettles views of parties as monolithic entities. Next, we provide several examples of intra-partisan gerrymandering. These span the one-party Democratic South, major Democratic cities, and heavily Republican states. We then argue that intra-partisan gerrymandering should be recognized as a distinct legal theory, especially under state constitutions. Its representational harms are analogous to those inflicted by partisan gerrymandering, and it can be regulated through a similar doctrinal framework. Finally, we offer an empirical proof of concept using a recent state senate redistricting as a case study. It’s feasible to identify intra-party factions, measure disparities in their treatment, and

The Congressional Record has preserved American legislative discourse since 1873, yet no unified platform exists for systematic analysis of this vast archive. We introduce CORA, the Congressional Oratory Research Archive: a comprehensive dataset spanning 1873 to 2025 with over fifteen million speeches linked to rich metadata including speaker identity, chamber, party, gender, state, date, and legislative references. Each speech connects to its original source through stable URLs, enabling verification against official proceedings. We enriched the corpus with topic labels derived from large language models and validated by human reviewers, supporting consistent analysis of issue attention across historical periods. CORA addresses gaps in existing resources by unifying historical and contemporary records, standardizing metadata, and extracting multiple resolution types beyond bills alone. The platform provides both bulk downloads in JSONL and CSV formats and an interactive web interface with search, filtering, and visualization tools. We validate CORA through conceptual replications of published findings on evidence-based rhetoric and gender-linked language, demonstrating that the corpus reproduces established linguistic patterns while extending coverage across 150 years of congressional history. By lowering barriers to access and analysis, CORA enables systematic study of how political attention shapes institutional change over time.

Conjoint survey designs are spreading across the social sciences due to their unusual capacity to estimate many causal effects from a single randomized experiment. Unfortunately, by their ability to mirror complicated real‐world choices, these designs often generate substantial measurement error and thus bias. We replicate both the data collection and analysis from eight prominent conjoint studies, all of which closely reproduce published results, and reveal high levels of measurement error in all. We then discover a common empirical pattern in how measurement error appears in conjoint studies and, with it, introduce an easy‐to‐use statistical method to correct the bias. Along the way, we provide a much simpler and simultaneously more powerful approach to designing, organizing, understanding, and analyzing conjoint data analyses.

Social media platforms increasingly mediate political information exposure, yet the role of algorithmic curation in shaping political exposure remains contested ^1,2 . This question is difficult to resolve on platforms in which users retain substantial control over their feeds ^3,4 . The ‘For You’ feed of TikTok, which delivers content almost entirely through algorithmic recommendation, offers a setting in which user agency is sharply constrained. Here we show, through 323 audit experiments with controlled ‘sock puppet’ accounts seeded with Democratic or Republican content across three US states, that accounts seeded with partisan content exhibited systematic, asymmetric differences in partisan exposure. Across more than 280,000 recommendations collected over 27 weeks during the 2024 US presidential election campaign, Republican-seeded accounts received about 11.5% more co-partisan content than Democratic-seeded accounts, whereas Democratic-seeded accounts were exposed to about 7.5% more cross-partisan content—largely anti-Democratic material—even after adjusting for engagement metrics. These asymmetries are concentrated among high-reach Republican channels and in specific policy domains, including immigration, crime and foreign policy for Democrats, and abortion for Republicans. Our findings show partisan imbalances in political information exposure on a platform dominated by algorithmic recommendations, with implications for platform governance and democratic discourse.

With the advent of remote-video technology and recent pushes to include video feeds in U.S. Supreme Court hearings, many are concerned about the effect that video and streaming might have on the behavior of U.S. judges and court participants. Previous research has shown that judges react to public sentiment, and anecdotal evidence suggests that introducing video recording might induce greater levels of “performative judging,” where behavior changes when there are cameras in the courtroom. We use AI-powered diarization to analyze the transcripts of oral arguments, leveraging the quasi-random adoption of video feeds in the U.S. Ninth Circuit Court. We find suggestive evidence that judges speak more and are more likely to interrupt attorneys when proceedings are video recorded rather than only audio recorded. However, these results are mixed and model dependent. When our estimation strategies account for interactive effects or include only judges that spoke in both audio-only and video hearings, we observe a camera effect for some outcomes of interest. By contrast, non-interactive models that include all judges result in null findings. These mixed findings provide novel but preliminary evidence exploring the performative judging hypothesis and suggest that courts and policy makers should consider the potential effects on judges’ behavior when deciding whether to introduce cameras into the courtroom.

Provides tools for analyzing data generated from conjoint survey experiments, a method widely used in the social sciences for studying multidimensional preferences. The package implements estimation of marginal means (MMs) and average marginal component effects (AMCEs), with corrections for measurement error. Methods include profile-level and choice-level estimators, bias correction using intra-respondent reliability (IRR), and visualization utilities. For details on the methodology, see Clayton, Horiuchi, Kaufman, King, and Komisarchik (2025) .

The United States government spends hundreds of billions of dollars annually obtaining goods and services from commercial bidders, a process known as government procurement. Without transparency, this process is susceptible to manipulation, corruption, and grift. However, the data required to assess the procurement process, including information detailing the amounts and awardees of government contracts, is difficult to obtain or study, though existing scholarship has made important contributions relying on limited samples of this data. We collect, clean, analyze, and make available the entire corpus of federal procurement contracts from 1979 to 2023, nearly 100 million contract actions over 45 years, and introduce an R package for accessing subsets of the data. Overall, these tools hold great promise for studying representation, corruption, and the connections between business and politics in the United States.