Polling after the 2026 midterms: Why the polls missed and how AI could change the game

TAMPA, Fla. - The final polls showed Republican Byron Donalds leading Democrat David Jolly by an average of four to five percentage points. Donalds won by just 1.3 points.

FOX 13's AI-assisted modeling anticipated a much closer race. On Election Day, FOX 13 Political Editor Craig Patrick described what the modeling was showing: "The tracking, the polling, the modeling, the latest returns, don't be surprised if this race winds up being 47-point something percent to 48-point something percent with Donalds being the victor."

Florida's governor's race: How close did AI get?

What we know:

Behind that assessment were multiple AI modeling runs incorporating FOX 13's own turnout and independent-voter calculations, pre-election polling on party loyalty and crossover voting, and other data. The most likely scenario projected a Donalds victory by 1.1 percentage points. The actual margin was 1.3 points — a difference of just two-tenths of a percentage point. The initial model also projected slightly more voters splitting away from the two major-party candidates to support minor-party and other candidates than actually did.

Veteran pollster Matt Towery, who is retiring from the polling business, believes artificial intelligence could eventually change how pollsters measure voters.

"Once the polling world decides to get its hands around, how do you harness AI and have it represent the way people think, rather than having to ask what people think? You can probably predict, create some pretty predictable models that'll be pretty much on target," Towery said.

AI modeling is not the same as traditional polling. Polls measure responses from people who participate in surveys. Models analyze available information and estimate outcomes based on their inputs and assumptions. And a close result in one election does not establish that a model will consistently outperform traditional polling.

Why the polls missed in 2026

The University of North Florida and St. Pete Polls were among the top-performing pollsters in Florida's governor and Senate races. But across the country, many surveys underestimated Democratic support. One of the biggest challenges was determining who would actually vote — and how motivated different groups were to participate. Democratic voters were more interested in the election, more politically engaged and more likely to turn out than many pollsters anticipated.

Exit polls also showed some voters who supported Donald Trump in 2024 shifting toward Democrats in 2026. Polls captured some of that movement but underestimated its magnitude in several states, including Georgia and Texas.

Towery believes pollsters may have overcorrected after struggling to capture Republican support in previous elections.

"Republicans were hard to get. In this cycle, to get them to answer. We thought it was because of the typical we're afraid to talk. In the end, we should have taken it one step further and figured out that we don't care. And that's really what the answer was," Towery said.

He pointed to Republicans who voted for Trump in 2024 but did not return to the polls in 2026.

"A lot of us got the fact that this was a Democratic wave. The question is whether we got that it was a tsunami," he said.

Towery's explanation highlights a central problem for pollsters: Getting a representative sample means more than finding people who will answer questions. It also means accurately determining which voters will show up.

How polling works — and why it is getting harder

By the numbers:

Florida has nearly 13.7 million registered voters. A typical statewide poll interviews up to 1,000 people. That means the odds of any one registered voter being selected for that particular poll would be about 1 in 13,700, assuming a simple random sample. For comparison, you're nearly 14 times more likely to win Florida's Pick 3 with a $1 straight play than to be selected for that hypothetical poll.

But that's the point of scientific sampling. Pollsters don't need to interview everybody. They need a carefully selected sample that reasonably represents the larger electorate. Consider a container filled with 13.7 million marbles. If 55% are red and 45% are blue, randomly drawing 1,000 should produce a sample reasonably close to that overall distribution.

The same principle applies to voters.

Selecting respondents is only part of the process. Pollsters must also estimate who is likely to vote. They may consider previous voting history, interest in the election, stated intention to vote, demographic characteristics, and political engagement. Those factors help pollsters decide which respondents to include in their likely-voter samples. And they may have to contact 10,000 people to get the 1,000 completed interviews they need.

The problem with reaching voters

Traditional telephone polling has become more difficult as landline use has declined. The people who still use landlines tend to be older and may not represent a cross-section of today's electorate.

"You're primarily reaching senior voters and your universe is shrinking every day," Towery said.

Cell phones create a different problem. People rarely answer calls from numbers they don't recognize. And those who agree to participate in telephone surveys may be more politically engaged than the average voter. There are also restrictions on automated calls to cell phones. Federal law generally requires prior express consent for autodialed calls or calls using prerecorded or artificial voices to wireless numbers. Manually dialed calls by live callers are permitted under that law.

Even when pollsters reach voters, getting honest answers can be difficult. Towery says voters in some states are particularly reluctant to discuss politics with strangers.

"Places like New Hampshire and Vermont and Iowa, some of these Midwestern states, they really don't like talking about their politics to strangers," he said.

He believes that reluctance contributed to polling misses involving Trump supporters in 2016, 2020 and 2024.

"You know, people didn't want to say they were voting for Trump. They felt like somehow that they were gonna be on a target list, people come over and they're dox them or get them, not that that's true, but that's how people think," Towery said.

Pollsters responded by making adjustments to ensure Republican voters were adequately represented in their samples. But in 2026, those adjustments may have gone too far. Pollsters placed considerable emphasis on capturing Republican voters they previously missed, while failing to account for changes in turnout and political engagement.

How pollsters are changing their methods

The backstory:

The industry is increasingly using a mix of approaches to reach voters, including survey panels and text messages. Panels consist of people recruited by polling companies to participate in surveys. Some participants receive compensation. Towery says that can create a problem: People who regularly take political surveys may be more interested in politics than the average voter.

What they're saying:

"Panels are created by large companies that find individuals who are willing to take polls. And these people are made aware of the fact that they will be asked to take surveys. In most of these instances, they are compensated," Towery said.

He believes that can produce a group that is unusually engaged and knowledgeable about politics.

"That isn't always the best way to do things," he said.

Not all panels are alike. Some use probability-based recruitment, while others rely on people who volunteer to participate.

Why text messages are gaining ground

Pollsters are also relying more on text messages to reach people who won't answer their phones. Towery previously told FOX 13 that the way questions are presented can make a difference in whether people respond. Now that he's retiring, he's willing to share more about his approach.

"People are very, very happy to respond to a text poll, as long as they're not hitting a link. Because once they hit that link, they don't know if it's, you know, what foreign country or somebody trying to get into their bank account," he said.

Instead, the question appears directly in the text conversation, and respondents scroll through the choices and select an answer. Towery says keeping the process simple is important.

He also believes shorter surveys can help reach people who aren't particularly interested in politics.

"Most pollsters, if you look at these polls that come in, they've got 60 questions on them. Now, anyone who's going to take a 60-question poll is someone who's really interested in politics," he said.

"But we really don't want people who are really interested in politics to respond. We want the person who's living their normal life."

Could AI change the future of polling?

Big picture view:

The polling industry faces a fundamental challenge: finding ways to measure an electorate that is constantly changing. Younger voters, Hispanic voters and other parts of Trump's 2024 coalition will be important to watch. Pollsters will need to determine whether the political shifts observed in 2026 continue or reverse. Artificial intelligence could offer another way to analyze those changes.

Rather than relying exclusively on surveys asking people how they intend to vote, AI could help identify patterns in voter behavior, turnout and political preferences. But the technology has limitations. Its conclusions depend on the quality of the underlying data and the assumptions built into the models. FOX 13's Florida governor's race analysis offers one example of the potential. But one election, however close the model came to the actual result, is not enough to establish a consistent track record.

What's next:

The next presidential election is already generating early polls. But until voters have a reason to pay attention to the candidates, campaigns and issues, those numbers tell us relatively little about how the election will turn out. Early polls can tell us where a race starts. They cannot tell us where it will finish. And for pollsters, the challenge remains finding better ways to reach voters, understand their priorities and accurately measure who will actually show up.