A Voter Pool Built From Social Media
A research team led by Fudan University, publishing under the name ElectionSim, says it collected more than 171 million posts from over 9.5 million X users during the 2020 election period, according to the team's paper, "ElectionSim: Massive Population Election Simulation Powered by Large Language Model Driven Agents." The researchers built classifiers to infer each user's age, gender, race, ideology and party affiliation from their public posting histories, then merged those inferred profiles with U.S. Census data and American National Election Studies survey data to construct a "million-level voter pool" capable of simulating presidential elections in every state. The paper reports that its latest version reproduced the winner in 47 of 51 state-level contests and 12 of 15 battleground states. The project's public page states the work is intended as a scientific exercise and disclaims any intent to influence actual U.S. elections. The system also lets researchers select modeled Americans by political or demographic traits and hold multi-round conversations with them, with a standing questionnaire covering topics including immigration, guns, race, LGBTQ rights, election integrity, taxes and civil unrest.
Testing A Message On Synthetic Pennsylvanians
According to reporting by journalist Natalie Winters, published on her Substack under the headline "EXCLUSIVE: China Is Building AI Models of American Voters. Why?," a separate project built roughly 300 simulated voter agents in each of ten states and used them to test political messaging. In Pennsylvania, researchers introduced a synthetic candidate who supported raising personal income taxes for wealth redistribution, then compared the modeled voters' reactions in treatment and control conditions across roughly 600 observations. Winters reports the treatment group expressed more concern, leading the study's authors to argue that voters' tax attitudes in the 2020 election may have been more conservative than commonly assumed. A longer questionnaire was then run with 100 Pennsylvania agents to probe why simulated Biden supporters with different education levels favored different policy framings; the investigation reports that college-educated agents emphasized language like "equity" and "fair," while non-college agents used different terms. Winters' reporting identifies the institutions behind this Pennsylvania work as SASS, described as a Chinese government academy, and SIIS, a think tank subordinate to the Shanghai municipal government.
A Broader Pattern Of Government-Linked Research
Winters' investigation, drawing on a review of Chinese-language research, university projects, datasets and government-linked think-tank publications, describes additional efforts beyond ElectionSim's academic modeling. It cites a Chinese study that constructed a "voter portrait of Trump's dependable right-wing base" and framed the work as offering a "scientific basis" for Chinese policy toward the United States, along with a state-funded follow-up project modeling "America First" attitudes to assess American positions on tariffs and alliances. Separately, according to Winters' related reporting on Chinese state researchers, a 2024 study from Tsinghua University's Center for International Security and Strategy examined how AI could shift political persuasion from broad demographic targeting toward individualized messaging based on a voter's inferred psychology and behavior, while a 2023 article sourced to the PLA National Defense University reportedly described data-driven "intelligent communication" in more explicit terms, and recommended individualized messaging from personal accounts as a way to avoid restrictions placed on official Chinese propaganda channels.
Context And Open Questions
The ElectionSim paper is an openly published academic work, and its authors explicitly disclaim any intent to influence U.S. elections, framing the project as scientific research into large-language-model-driven social simulation. The additional government-linked studies and the Pennsylvania messaging test described in Winters' investigation are reported through her review of Chinese-language publications rather than through documents she states were independently obtained from the institutions themselves, and it remains unclear whether any of this modeling has been used operationally rather than academically. The reporting sits alongside separate, previously documented Chinese influence activity: NPR and Graphika have reported on a long-running campaign known as "Spamouflage" that used fake accounts posing as American voters to push divisive content on U.S. political topics, and Microsoft warned in 2023 that suspected Chinese operatives used AI-generated images to mimic American voters online. Those campaigns involved fabricated personas posting directly on social platforms, a different mechanism from the survey-style voter simulation and message-testing described in the Fudan and government-linked research.