Synthetic Voters Aren’t Voters: Where AI Belongs in Campaign Research and Where It Doesn’t

Imagine a campaign manager waking up the morning after a debate and asking an AI model what undecided voters now think.

Within minutes, the model produces a convincing answer. It identifies which arguments resonated, which groups may have moved and which message the campaign should use next. I understand the attraction. Campaigns are always working under pressure, and this kind of information would normally require time, respondents and money.

The problem is not necessarily that the answer is wrong. The problem is that no voter actually gave it.

A synthetic respondent is not a faster or less expensive voter. It is a model-generated prediction of how a person or someone with a particular demographic profile might answer. That prediction may be useful, but it is fundamentally different from asking a human being what they believe.

Campaigns need to understand that distinction before plausible simulations quietly begin to replace evidence.

The Real Danger Is a Plausible Answer

If synthetic voters produced obviously ridiculous results, they would not present much of a problem. The greater risk is that their answers often sound entirely reasonable.

Large language models are extraordinarily good at recognizing patterns and creating coherent representations of them. Give a model the profile of a 52-year-old suburban Republican woman, and it can produce a convincing account of what that person might think.

But real voters are not simply the average of everything previously written about people who resemble them. They hold contradictory opinions, misunderstand policies and react emotionally. They change their minds for reasons that do not fit neatly into demographic categories. Two voters with nearly identical profiles may interpret the same political message in completely different ways.

Synthetic data can smooth over precisely the disagreement, uncertainty and unpredictability that campaigns most need to understand.

Pew Research Center summarized some of these concerns and found that AI-generated estimates can stereotype demographic groups, have greater difficulty representing Republican viewpoints than Democratic ones and understate the actual level of disagreement within public opinion. Pew itself interviews real people rather than using AI to decide what the public thinks. Its explanation of the methodological concerns is worth reading.

Academic research gives us further reason to be careful. One study compared synthetic responses with human data from the American National Election Studies. It found that 48 percent of the relationships estimated using AI-generated responses differed significantly from their human-derived equivalents. Among those differences, nearly a third pointed in the opposite direction.

The researchers also found that results changed when prompts were adjusted and when the same exercise was repeated after the underlying model had been updated.

That does not mean synthetic data never aligns with human data. A 2025 proof-of-concept study using Chilean public opinion data achieved excellent results on certain trust measures. At the same time, the accuracy varied substantially across questions and demographic groups. The researchers concluded that careful calibration and additional testing were still necessary.

This is an important distinction. Synthetic respondents are not automatically inaccurate. Their accuracy simply cannot be taken for granted.

Politics Is a Difficult Environment for Synthetic Certainty

In relatively stable areas of consumer research, historical patterns may remain relevant for some time. How people feel about a household product today may be very similar to how they felt six months ago.

Politics moves differently. A debate, court decision, economic report, international crisis or candidate mistake can change the context within hours. Unless a model is supplied with fresh, representative human data, it has no direct knowledge of how voters interpreted that event.

Access to current online content does not solve the problem either. Social media is not the electorate. It tends to amplify the people who are most active, emotional and visible. Those voices may be politically important, but they are not necessarily representative of the voters who will determine the outcome of a close race.

The most important voters may be the ones who depart from the established pattern. It could be a habitual non-voter newly motivated by an issue, a partisan reconsidering an old loyalty or an undecided voter reacting in a way that historical data did not anticipate.

A model may treat these people as statistical noise. A campaign cannot afford to do the same.

There Is Still an Important Role for AI

I am not arguing that AI has no place in political research. Used responsibly, it can make research faster, sharper and more useful.

AI can help identify ambiguous survey questions, test questionnaire logic, explore alternative wording and anticipate how a question might be misunderstood. It can help organize thousands of genuine open-ended responses, surface themes for human review and compare possible message hypotheses before those messages are tested with voters.

It can also help researchers work more effectively with information they have already collected. The dividing line should be clear: AI can support the process of understanding voters, but it should not silently become the voter.

Before using synthetic responses to inform a campaign decision, researchers should ask:

• What fresh human evidence anchors the model’s output?
• Has the result been validated against the relevant electorate?
• Does it preserve disagreement and variation within groups, or does it produce a convincing average?
• Would the conclusion remain stable if the prompt, model or model version changed?
• Does everyone using the findings understand which responses came from people and which were generated?

Transparency is particularly important. The American Association for Public Opinion Research now explicitly states that AI-generated responses are not research participants. It also says that terms such as “poll” and “survey” imply that the primary data came from human respondents. AAPOR requires synthetic cases to be clearly identified as AI-generated.

Campaigns should use AI to improve human research, not to remove humans from it.

Synthetic voters can help us formulate better questions, challenge assumptions and explore possibilities. They cannot experience rising prices, watch a debate, lose trust in a candidate, feel newly represented by an issue or decide to stay home on Election Day.

When the question is what voters believe, there is still no substitute for asking voters.

Allen Porter is General Manager of Enghouse Insights.