待翻譯：China Is Building AI Models of American Voters

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Fudan built a million-account X “voter pool.” Researchers from Chinese government think tanks tested synthetic Pennsylvania voters with a campaign message. A state-funded project modeled “America First” attitudes to assess tariffs, alliances, and U.S. foreign policy. China is building increasingly detailed models of the American electorate—who voters are, what they believe, why they support Donald Trump or Joe Biden, and how they might respond to political messages. A review of Chinese-language research, university projects, datasets, and government-linked think-tank publications reveals that Chinese institutions have moved far beyond simply following American polls. Researchers are now constructing artificial versions of U.S. voters, assigning political and demographic labels to more than one million X users, simulating state-by-state presidential elections, and conducting experiments on synthetic voters in Pennsylvania. Another Chinese study, funded by the country’s National Social Science Fund, used American polling data to infer “America First” attitudes toward tariffs, alliances, international organizations, and democracy promotion. Its author explicitly argued that the research could help China evaluate the future direction of U.S. foreign policy and formulate a response. Chinese universities and government think tanks are developing increasingly granular, interactive models of the American electorate that can be used to forecast political behavior, test messages, interview simulated constituencies, and identify the domestic forces shaping U.S. policy toward China. Their own documents raise a simple question: Why? A Million-Account American “Voter Pool” The most expansive project uncovered in this investigation began with an enormous collection of posts from X, formerly Twitter. In a paper titled “ElectionSim: Massive Population Election Simulation Powered by Large Language Model Driven Agents,” a Fudan University-led research team said it collected 171,210,066 posts from 9,596,198 X users during the 2020 election period. Researchers developed classifiers to infer users’ age, gender, race, ideology, and party affiliation from their public histories. They then combined those profiles with Census and American National Election Studies data to simulate presidential elections in every state. The latest version of the paper claims the system reproduced the winner in 47 of 51 state-level contests and 12 of 15 battlegrounds. The project also allows users to select modeled voters by political or demographic traits and conduct multi-round conversations with them—effectively creating synthetic focus groups derived from real social-media behavior. They also ran a mock 2024 election. The researchers’ questionnaire goes far beyond candidate preference. It includes immigration, firearms and defense, race, LGBTQ issues, gender resentment, and democratic norms. ElectionSim subsequently fed into a broader Fudan-led system called SocioVerse, described by its creators as a “world model” powered by a pool of more than ten million real-world social-media users. Synthetic Pennsylvania Voters Get a Campaign Message A separate research project went one step further: it used artificial voters to conduct a political-message experiment in a critical swing state. The paper, “Intelligent Computing Social Modeling and Methodological Innovations in Political Science in the Era of Large Language Models,” was written by researchers affiliated with the Shanghai Academy of Social Sciences, Shanghai Institutes for International Studies, Nanjing University, and Shanghai Jiao Tong University. Two of those affiliations are especially important. The Shanghai Academy of Social Sciences is a government academy. The Shanghai Institutes for International Studies, or SIIS, describes itself as a high-level research institution subordinate to the Shanghai municipal government whose mission is to provide intellectual support for party and government decision-making. Share The researchers proposed a framework called “Intelligent Computing Social Modeling,” or ICSM. Using 2019 American Community Survey distributions, they constructed artificial voters with combinations of ethnicity, gender, age, region, education, occupation, and industry. The study simulated voters in Pennsylvania, Ohio, Michigan, Missouri, Indiana, West Virginia, California, Texas, Wisconsin, and Georgia. Researchers deployed 300 agents per state and prompted the models to choose between Democratic and Republican candidates. Then they conducted a list experiment on 300 Pennsylvania agents. The treatment introduced a candidate who supported raising personal income taxes for wealth redistribution. By placing the same synthetic profiles into treatment and control conditions through separate model calls, the researchers generated 600 observations and measured the change in stated concerns. The treatment group reported more concerns, leading the authors to argue that voters’ tax attitudes in the 2020 election may have been more conservative than commonly understood. The researchers also conducted a longer questionnaire with 100 Pennsylvania agents. They examined why simulated Biden supporters with different education levels favored particular policies. According to the paper, college-educated agents emphasized words such as “equity” and “fair,” while non-college agents placed greater emphasis on issues including the minimum wage. Modeling Trump’s Base—for China’s U.S. Policy Other Chinese research documents are clearer about why modeling American voters is useful to China. In 2024, Shanghai Jiao Tong University scholar Shu Fu published a study titled “Who Supports Trump? An Empirical Analysis of American Populism.” Using American National Election Studies surveys from 2012, 2016, and 2020, Fu measured political, economic, and identity-based anxieties and used multiple statistical models to explain support for Trump. Leave a comment The paper said it sought to: “勾勒出支援特朗普的右翼鐵票的選民畫像” Or: “sketch a voter portrait of Trump’s dependable right-wing base.” Fu concluded that identity anxiety—particularly views concerning immigration and the status of white Americans—was the strongest predictor of support for Trump. But the paper explicitly placed its findings in a Chinese policy context. It said understanding the voters behind Trump could help explain the mechanisms of American domestic politics and “provide a scientific basis for our country’s policy toward the United States.” Fu subsequently expanded the work in a 2025 paper titled “The Long-Termization of ‘America First’: The Public-Opinion Foundation of Populist Foreign Policy.” The methodology joined two major American polling resources. Fu first used 2020 ANES data to calculate political, economic, and cultural populism. Because ANES contains limited foreign-policy questions, Fu trained predictive models using variables shared with the Chicago Council on Global Affairs’ 2022 survey—including vote choice, ideology, party identification, religion, education, income, race, age, sex, marital status, and homeownership. The models were then transferred into the Chicago Council data, producing inferred political, economic, and cultural populism scores for each anonymized respondent. Those scores were used to estimate relationships between American populism and support for tariffs, isolationism, alliances, international organizations, and the promotion of democracy abroad. The study concluded that political populism increased suspicion of alliances and international organizations; economic populism increased support for tariffs; and cultural populism increased opposition to democracy promotion and pluralist values. The paper framed those findings as useful for evaluating the future course of U.S. foreign policy and China’s response. Its replication data and code were also deposited in Harvard Dataverse. Thousands of American Survey Respondents Become AI Voters The technique is spreading across Chinese universities. A Wuhan University paper converted the demographic characteristics of 6,571 respondents in the 2020 ANES survey into large-language-model personas. The model was instructed to adopt the identity of each anonymized respondent and cast votes in the 2016 and 2020 elections. Researchers then aggregated the artificial votes by state. Before the 2024 election, they used the same personas to forecast a contest between Trump and Kamala Harris, predicting a Trump victory and roughly 300 electoral votes. AI As Political Warfare Internal documents from the Chinese technology company GoLaxy, collected in a Vanderbilt University archive, describe a system intended to serve national-level “cognitive-domain confrontation and guidance.” The company claimed portions of its technology had been deployed to the Ministry of State Security, PLA Joint Staff intelligence elements, and Strategic Support Force units. The documents include a U.S. election module featuring political-influencer rankings, follower profiles, polling, betting markets, Federal Election Commission data, swing-state history, and “supporter portraits.” Other screens show MAGA-related content and organization libraries using the National Rifle Association and National Endowment for Democracy as examples. The question is no longer whether Chinese institutions are building models of American voters. It is what they intend to do with them. Share Leave a comment Share Natalie Winters