Selected Manuscripts
Selected Manuscripts
AI and Political Information
Why Do AI Models Tell Left-Wing Voters to Support the Communist Party? AI Voting Advice in Japan’s 2026 General Election
Miyazaki, Sho and Andrew B. Hall.
Abstract
AI chatbots with web search are increasingly used for voting guidance, yet we know relatively little about their recommendation behavior outside the US. We query five models from three companies with 36,300 synthetic voter profiles during Japan’s 2026 Lower House election, plus a follow-up with today’s four frontier models. Policy stances dominate recommendations, producing 50–98 percentage point swings compared to 0.5–7.0 pp for demographics. Strikingly, left-leaning stances cause all five models to converge on Japan’s Communist Party, despite other parties holding similar positions on the relevant issues. We trace this to the information environment: JCP operates a fully open website with a party newspaper that models freely access and frequently misclassify as independent journalism, while major Japanese news outlets block AI crawlers. Without policy input, models show no uniform left-wing bias. Incorporating X search shifts recommendations modestly leftward, contradicting US-centric expectations. These findings suggest that AI models need better source discrimination in non-US contexts, and that news organizations blocking AI access may inadvertently cede influence over AI-mediated voting advice to partisan sources that remain open.
Newspapers Blocking AI Crawlers Expose Their Voters to a Partisan Flood
Miyazaki, Sho.
Abstract
Major newspapers block AI crawlers, while parties and partisan sites leave their text open. This paper builds a game-theoretic model of the hold-out and asks what voters hear while it lasts. In the model, an AI that cannot grade its sources pools the text it can read; two persuaders choose how much nearly free text to publish; and an AI lab may license a newspaper’s reporting for its lost revenue. Neither the newspapers nor the lab has a political aim. Three results follow. First, the persuaders flood the AI, and the lab licenses each newspaper once a fixed share of the voters closest to it asks the AI, whatever the AI says. Second, newspapers whose voters dislike the AI’s answers more hold out longer, and their voters hear the flood longer. Third, licensing bends the persuaders’ contest by a small tilt from where the costs alone put it: patient persuaders tilt the AI’s answers against the voters’ average ideology, and impatient ones toward it. Business decisions with no political aim thus set how long each group hears the flood and which way the AI’s answers tilt.Online Voting Systems
What Happens When Anyone Can Be Your Representative? Studying the Use of Liquid Democracy for High-Stakes Decisions in Online Platforms
Hall, Andrew B. and Sho Miyazaki.
- Non-technical summaries available at the a16zcrypto article and the Hoover Institution video.
Abstract
Since the 19th century, political reformers have proposed broadening civic and corporate governance by allowing voters to delegate to any other voter—sometimes known as liquid democracy. Today, systems like liquid democracy have become an important part of ongoing efforts to create democratic online platforms governed by users rather than elites. We provide a first empirical political science study of liquid democracy in a high-stakes, real-world setting, analyzing data from over 250,000 voters and 1,700 proposals across 18 crypto projects (“DAOs”) built on the Ethereum blockchain. We find that, on average, 17% of voting tokens are delegated, with substantial clumping on the most-popular delegates. Delegation is primarily bottom-up, with smaller token-holders more likely to delegate. More active voters receive more delegations, suggesting somewhat informed decision-making. Using a difference-in-differences design, we estimate that creating online hubs to coordinate delegation significantly increases delegation and overall voting rates. In sum, liquid democracy can foster bottom-up participation, particularly when paired with tools for coordination. On the other hand, real-world participation remains relatively low among both voters and delegates, posing an important challenge to liquid democracy not yet contemplated in existing theoretical literature.How Does a Public Tally Affect When Voters Vote? Evidence from Online Roll Call Voting
Miyazaki, Sho.
Abstract
Do voters ‘’signal’’ their decisions by voting early, or do they ‘’snipe’’ by casting their votes at the last minute? In the transparent sequential roll call voting of the U.S. House of Representatives, members can strategically choose when to cast their votes, but the timing of those votes is hard to observe and analyze. This paper examines the effect of a public tally on vote timing in a setting that records it exactly: corporate shareholder voting at a blockchain-based startup (DAO) whose roll call votes transitioned from private one-stage voting to transparent sequential voting. I collected individual-level voting records, delegated voting power, and precise vote timing. Fixed-effect regressions show that introducing a public tally increases the likelihood of early voting (signaling) or last-minute voting (sniping), consistent with game-theoretic predictions. Interaction estimates are consistent with more pivotal voters, measured by the Shapley-Shubik index, using early votes as signals rather than waiting to snipe. These results suggest that a public tally may harm information aggregation, as signaling can intensify the influence of pivotal voters.Redistricting and Gerrymandering
Malapportionment as a Gerrymandering Strategy: Exploring Electoral Manipulation in Malaysia Using a Simulation Algorithm
Miyazaki, Sho and Yuko Kasuya.