What Campaigns Need To Know About AI Polling

AI-generated polling has made the leap from a what-if to a reality. Startups are already selling synthetic surveys to campaigns, PACs, and corporations. Trusted brands like Gallup and Ipsos are running their own studies to validate the concept. We have new rigorous, publicly available evidence about how well the technology performs.

This article summarizes what we know today. What is synthetic polling? What does the research say about its capabilities? What has to happen before campaigns can rely on it for real strategic decisions in the same ways they turn to traditional polling?

The Problem With Polling Today

Traditional polling – where we ask real individuals for their opinions via phone, text, or web – is getting more expensive and less reliable because of the response rate challenge. Just 0.5 to 2 percent of Americans will respond to a survey request. In order to achieve a representative sample, we now have to field surveys longer (impacting their accuracy) and reach more voters (increasing the cost). For a statewide campaign, a single high-quality poll can run $30,000 to $50,000 and down-ballot campaigns may be priced out entirely from quality survey research.

At the same time, large language models have gotten better at mimicking how real people write and speak, so it's not a crazy leap to wonder whether AI can also answer questions like a voter would. That's the thesis behind a new survey methodology powered by AI called synthetic sampling, silicon sampling, or digital twins. Instead of interviewing humans (and the cost associated with doing so), you build simulated respondents and ask the AI model to predict how each one would answer the questionnaire.

Aaru, a startup founded in 2024, reached a $1 billion valuation after its AI agents predicted a New York Democratic primary within 371 votes, and it now sells polling to campaigns, super PACs, and Fortune 500 companies at less than a tenth the cost of human surveys. Gallup announced in May that it's testing simulated respondents with Simile, a company founded by the Stanford researchers who pioneered the method. Ipsos has its own Stanford partnership. And on the open-source side, projects like MiroFish – a multi-agent simulation engine that topped GitHub's trending list this spring – let anyone spin up thousands of AI agents that read the news, argue with each other on simulated social platforms, and produce a prediction report. The technology is not hypothetical, and it is not going away.

How AI Polling Compares

The most rigorous public test of synthetic polling comes from Verasight, a survey research firm, which published their findings in a June report. They built and iterated on a robust synthetic polling framework built on real data and interviews. They compared the AI's output to real surveys on a set of questions. Verasight's CEO, Ben Leff joined me on a recent episode of the Campaign Trend Podcast to discuss the study.

Three results from that research stand out for campaign decision makers.

First, synthetic samples are actually good at answering the questions we least need help with. On Trump approval and the generic ballot, the AI's error was 1.8 to 2.1 points, roughly what you'd expect if you fielded the same poll twice. But those are the most heavily polled questions in America and it's not difficult for a human to predict that a registered Democrat who voted for Kamala Harris in 2024 doesn't approve of the president right now.

Second, the errors show up exactly where campaigns need precision. On questions involving recent events, policy comparisons, or any topic without a strong partisan signal, errors ran from 8 to nearly 13 points. And we can't tell in advance which questions the AI model will nail and which it will miss. Crosstabs – where we look at specific subgroups – are even more challenging. Synthetic Republicans, for example, weren't conservative enough nor were synthetic Democrats liberal enough. So if you're trying to use a poll to decide which message moves suburban women or younger Black men, that's disqualifying. To make matters worse, the model didn't know how to handle breaking news that occurred after its training, which is exactly when campaigns need fresh polling.

Third, and most surprising, is that throwing more data at the problem doesn't reliably help. Verasight found that adding voter file history, more elaborate prompts, and newer models sometimes improved accuracy on one question while making another worse. Feeding current news to the agents also backfired. Even scaling from 500 to 5,000 AI respondents barely moved the error. Unlike with traditional polls, these aren't sampling errors but endemic to the model.

A 2024 study found that AI responses tend to compress the natural variation of human opinion, shift with minor changes in prompt wording, and drift over time, meaning the same prompt produces meaningfully different results.

Are We Judging AI Polls The Right Way?

Every evaluation of AI polling so far benchmarks against human polls. That seems like a reasonable starting point because that's what we currently use in campaigns, but it may be the wrong yardstick, especially as fragmentation of media challenges the underlying concept of representative samples polling is built upon. A poll is a measurement of what people said. A synthetic sample is a prediction of what people would say. One way to think about it is the difference between a weather forecast and the actual observation at a given moment. Judging a prediction by how well it imitates a measurement tells you something, but that's not the point.

The better question for a campaign decision maker is: does the tool improve how I allocate resources and direct strategy? Despite what reporters, pundits, and other observers think, a poll is not an end in itself. It is one input to decisions like where to spend, which messages to run, and which voters to target. If an AI poll moves those decisions in the right direction faster or cheaper than the alternative (which in some cases is simply gut instinct) then it should earn its place in the campaign toolkit even if it fails in a head-to-head against a traditional benchmark survey. But, if it moves those decisions in the wrong direction, it doesn't matter how cheap or quick it is.

The traditional survey itself is even becoming less reliable. When Mr. Gallup's survey researchers were going door to door, the whole country was still watching Cronkite on the evening news and reading the same local newspaper. Everyone being polled roughly had the same limited inputs so we could reasonably expect two demographically similar individuals to share similar opinions. Today, our media diets are as unique as our DNA. In other words, the gold standard of live human polling may be tarnishing.

What Campaigns Can Do With AI Polling Now

Based on current capabilities, there are a few narrow, practical use cases for synthetic sampling. Campaigners can pre-test their surveys before putting them into the field. An AI poll might flag confusing wording or give you an initial read on whether an issue is 50-50 or 80-20. This information can lead to cost savings by increasing completion rates or trimming questions.

Directional reads where being roughly right is acceptable are another area where AI polls can help campaigns, especially if they are otherwise priced out from traditional survey methods. If the choice is no data or synthetic data, some information is better.

At this point, you shouldn't treat AI crosstabs like real crosstabs, use synthetic data to test messages on subgroups, or practice synthetic boosting where a real sample's hard-to-reach demographics are padded with AI respondents.

If you're evaluating an AI polling product, three questions will tell you most of what you need to know.

1. Ask to see the synthetic predictions side by side with a real survey on the same questions, fielded at the same time – and not just presidential approval, where every model looks good.
2. Ask to see the crosstabs, because topline numbers can look right while the subgroups underneath are badly wrong.
3. And ask how the model handles events from the past month.

Vague answers to any of these tell you what you're buying.

The AI Polling Outlook

Two things stand between today's AI polling technology and a decision making tool campaigns can actually rely on.

Campaign decision makers thrive on pattern recognition. We simply don't have the reps under our belts to know when synthetic numbers are trustworthy or not. That can only come with time, in real campaigns where AI polling is in the toolbox. This is the very beginning of the learning curve for synthetic sampling and there are no shortcuts. It's worth understanding the economics here: the cheap part of an AI poll is the computing, but the ingredient that makes it credible is real human survey data. Verasight's pipeline required roughly 15,000 real interviews to work at all, which means synthetic polling is downstream of traditional polling, not a replacement for it.

We need to figure out what data actually improves the AI polling models. It's not simply a matter of adding more data, but adding the right data. Early studies show that gains come from using real opinion data and hybrid approaches that blend statistical models with AI responses. The cheap part of an AI poll is the compute, but the ingredient that makes it credible is real human survey data. We're still in the kitchen testing out recipes.

Conclusion

AI polling is coming to campaigns. The cost pressure is too strong and the technology is too alluring. The history of AI is recent enough that we can remember what ChatGPT 3.5 was like. "This is the worst AI will ever be" was a constant refrain. Right now, synthetic sampling isn't ready to supplant traditional polling.

But we can't afford to dismiss the technology outright. We have to keep experimenting with the tools, using them where we can, and assess what AI polls actually tell us.