China Is Quietly Building AI Copies of American Voters — And Testing Political Messages on Them

.

A New Kind of Election Research

Imagine a computer program that can act like a real American voter. It knows how that person might vote, what worries them, and how they might react to a campaign message. Now imagine millions of these fake voters, built from real online activity.

That is exactly what several Chinese university teams have created in recent years. Researchers used artificial intelligence (AI, computer systems that can learn and make predictions) to build detailed digital models of the U.S. electorate.

These are not secret spy tools hidden in a basement. Much of the work has been published in open, publicly available academic papers. But the scale and detail involved have caught the attention of American commentators, lawmakers, and security researchers.

Millions of Fake Voters, Built From Real Posts

The most striking project comes from Fudan University in Shanghai, one of China's top research schools. A team there built a system called ElectionSim.

According to the team's own published paper, researchers collected more than 171 million posts from over 9.5 million users on the platform then known as Twitter (now X), all gathered during the 2020 U.S. election period. The team built the tool to model voter preferences in specific election scenarios, collecting the huge dataset to construct a diverse voter pool with over a million distinct users.

Using this data, the researchers built computer profiles guessing each person's age, gender, ethnicity, political leaning, and party. They then combined these guesses with U.S. Census data and election-survey data to simulate how entire states might vote.

The results were unsettling in their accuracy. In one independent academic review of the tool, ElectionSim's simulation of the 2024 race projected Kamala Harris and the Democrats winning 8 of 15 battleground states, with Donald Trump and the Republicans projected to win the remaining 7 — a genuinely close result, much like real pre-election polling.

The system does not stop at predictions. Its creators built the tool so that a user can pick out simulated Americans by traits like political leaning or background and hold back-and-forth "conversations" with them, testing how these fake voters might respond to arguments or news events. The project's full code is openly available; the team behind it credits support from Fudan University, Indiana University, and the University of Rochester.

Testing a Tax Message on Simulated Pennsylvania Voters

A second, related research effort went even further. It did not just simulate voters — it experimented on them.

A paper from scholars connected to the Shanghai Academy of Social Sciences (a government-run academic body), the Shanghai Institutes for International Studies, Nanjing University, and Shanghai Jiao Tong University built artificial voter profiles for ten U.S. states, including the key battleground of Pennsylvania.

The researchers then introduced a fictional political candidate who supported raising personal income taxes to redistribute wealth, and measured how the simulated voters reacted before and after hearing that message. This is a technique political consultants use in the real world to test slogans — except here, it was run entirely on computer-generated fake Americans rather than real focus groups.

Some Studies Say the Goal Is Understanding Trump's Base — For Chinese Policy

Not every paper hides its purpose. Some are explicit about why understanding American voters matters to Beijing.

A 2024 study by a scholar at Shanghai Jiao Tong University analyzed what drives support for President Trump, using long-running American academic election surveys. The paper concluded that anxiety over identity and immigration was the strongest predictor of support for the populist right.

Crucially, the author stated the goal was to help explain U.S. domestic politics and to provide, in the researcher's own words, a scientific basis for shaping China's policy toward the United States. A follow-up 2025 paper by the same researcher went further, building statistical models to estimate how populist attitudes among Americans connect to support for tariffs, skepticism of alliances, and opposition to promoting democracy abroad — explicitly framing this as useful for predicting the future direction of U.S. foreign policy.

Not an Isolated Case

This is not the first time Chinese-linked technology aimed at understanding or influencing American politics has surfaced. A separate and larger case, unrelated to the university papers above, gained attention in 2025.

Documents examined by security researchers showed a Chinese company called GoLaxy had built detailed data profiles on at least 117 sitting U.S. lawmakers and more than 2,000 other American political and public figures, as part of technology researchers described as built for large-scale, AI-driven influence operations. The firm was founded in 2010 by a research institute connected to the state-run Chinese Academy of Sciences.

U.S. officials reportedly monitored the company during the 2024 election to check whether it was trying to shape American political conversations, though they concluded at the time that it was not actively doing so. The case still worried some U.S. lawmakers enough that they raised it directly with America's top intelligence official.

This broader pattern is worth watching in the context of China's own history with academic election forecasting. Fudan's Center for Complex Decision Analysis, a separate group from the ElectionSim team, has been running election prediction models publicly for years. Just two days before the 2024 U.S. election, that Fudan center publicly forecast Trump as more likely than not to defeat Harris, based on a computer simulation predicting vote shares across eight key states.

Why This Matters

None of this proves that Beijing is currently running a live influence campaign against American voters using these exact tools. Most of the research cited here was published openly in academic journals, and some of it, like the Fudan election forecast, was framed by its own authors as purely academic work with no intention of swaying the vote.

But the pattern is hard to ignore. Chinese state-linked and state-adjacent institutions have built, tested, and refined the technical ability to model how American voters think, to predict how they'll vote, and to measure how they respond to political messages — all with a particular focus on understanding the populist, pro-Trump wing of American politics. Whether these tools stay in academic journals or get put to more active use is, as one investigation into the topic put it, no longer the real question. The real question is what they're for.

Ordinary voters have no way to know if a post or account moved by one of these systems has ever crossed into their own feed. That uncertainty is itself part of the concern that lawmakers and researchers are now raising.

.

Sources

1. Fudan University, Indiana University, University of Rochester — "ElectionSim: Massive Population Election Simulation Powered by Large Language Model Driven Agents" (research paper, arXiv): https://arxiv.org/abs/2410.20746
2. ElectionSim project source code (GitHub): https://github.com/amazingljy1206/ElectionSim
3. South China Morning Post — "Researchers in China use computer modelling to predict Trump election victory": https://www.scmp.com/news/china/diplomacy/article/3285125/researchers-china-use-computer-modelling-predict-trump-election-victory
4. Nextgov/FCW — "Researchers detail new 'gray zone conflict' in AI-driven Chinese propaganda": https://www.nextgov.com/cybersecurity/2025/08/researchers-detail-new-gray-zone-conflict-ai-driven-chinese-propaganda/407358/
5. The Register — "Chinese biz using AI to influence US politicians": https://www.theregister.com/2025/08/08/golaxy_ai_influence/
6. GA-S³ paper (independent academic review citing ElectionSim's 2024 simulation results, arXiv): https://arxiv.org/pdf/2506.03532

.