Can AI Stand In for Voters? Pollsters Are Testing the Idea

Some polling firms are experimenting with surveys that use artificial intelligence personas instead of asking only real people what they think, according to reporting by Phys.org. The idea is called synthetic polling: researchers create hundreds or thousands of simulated people using large language models, or LLMs, which are AI systems trained to produce humanlike text. These personas are built from existing polling data and demographic information such as age, location, or background, then asked survey questions or placed in artificial focus groups.

The appeal is easy to see. Real polling takes time and money. A good national survey has to reach a group that is nationally representative, meaning it reflects the larger population closely enough that the results can be taken seriously. Jennifer Stromer-Galley, a Syracuse University professor who studies public opinion polling methods, told Phys.org that a strong national survey should include roughly 1,200 people. Synthetic polling firms, by contrast, can generate several thousand AI personas without having to persuade actual people to answer questions.

That creates the central trade-off. Synthetic polling could let researchers test questions quickly, explore policy ideas, or run focus groups at a lower cost. But the same technology may also create a misleading picture of what the public thinks. The worry is not just that an AI system could make a mistake. It is that the system may be too consistent in a way real people are not.

Public opinion changes because life changes. A person asked in October about the country’s biggest problem might give a different answer in January after a job loss, a storm, a new law, a family event, or simply months of news. Stromer-Galley’s concern is that many AI personas are built from fixed instructions and background information. If the model is asked the same kind of question again and again, it may give similar answers again and again. That consistency sounds useful, but it can hide the instability and contradiction that are part of human opinion.

Real voters also do not always think like neatly organized policy spreadsheets. A person might support one government program but oppose another similar one, or might care deeply about an issue only after it affects their community. Stromer-Galley argues that synthetic personas may produce answers that are more internally consistent than human answers. In polling, that can be a problem, because the point is not to create tidy opinions. The point is to measure the opinions people actually have.

Another issue is bias, which in this context means a built-in pattern that pushes results in one direction rather than reflecting the full range of views. Polling vendors can adjust their synthetic samples so the personas match the average demographics of a population. But Stromer-Galley says that does not solve the problem of how the AI model answers for each profile. If a model is asked to simulate a rural Black voter in a small town, for example, it may produce what seems like a typical answer for that profile. But real communities include outliers, people whose views do not fit the expected pattern. If the synthetic method smooths away those outliers, it may miss something politically or socially important.

A Pew Research Center experiment suggests this concern is not just theoretical. Pew had an AI model take on the personas of real members of its American Trends Panel and answer the same surveys those human panelists took. Across nearly 300 questions, the AI results differed from the human results by an average of 12 percentage points. Nearly half the questions had at least one answer choice that no AI respondent selected. That matters because a survey that never selects an answer real people sometimes choose is not simply noisy; it may be leaving out parts of public opinion.

The debate is also about what polling is for. Since scientific public opinion polling developed in the 1930s and 1940s, it has offered more than numbers. It has given people a way to see how others across the country think about shared problems. If synthetic polling becomes common, it could expand access to opinion research, especially for organizations that cannot afford large surveys. But it could also create what Stromer-Galley calls a distorted mirror: a reflection of the public that looks convincing while bending reality.

That does not mean AI personas have no place in research. They might help test survey wording, explore possible reactions, or supplement traditional methods. The harder question is where to draw the line. If synthetic polling is treated as a tool, with clear warnings about its limits, it may be useful. If it starts replacing real respondents, the public may end up being measured by machines trained to imitate people, rather than by people themselves.