Wisconsin and Michigan Just Ran the Experiment on Public Polling — and the Instrument Failed Exactly the way it was Built to Fail

Six days ago, in The Polling Paradox of 2026, I argued that the most-polled midterm in American history had produced the least trustworthy numbers in modern memory, and that public polling had stopped measuring the electorate and started measuring the polling industry. I assumed the argument would have to wait until November for a verdict.

It waited five days.

On Tuesday, August 11, three Democrat primaries in the upper Midwest ran the experiment in public, and the results are worse than my thesis required.

Wisconsin: twenty-one points to nothing

Start with the governor’s race. On July 29 the Marquette Law School Poll — the most respected survey operation in the state, run by a serious methodologist — put state Representative Francesca Hong at 38 percent, Mandela Barnes at 16, and Milwaukee County Executive David Crowley at 7, with 34 percent undecided. Push the leaners and Marquette had Hong at 46 percent of the primary electorate. A week later, State Navigate polled 1,473 likely Democratic primary voters between August 3 and August 6, with a stated margin of error of 2.3 points, and found Hong leading Crowley 44 to 24. Narrowed to a two-way question, Hong led 49 to 28.

Crowley won. He took 39.8 percent to Hong’s 39.3 percent — a margin of 3,783 votes out of nearly 790,000 cast, or half a point. The Associated Press could not call it until early Wednesday morning, and only did so after Hong conceded and removed the possibility of a recount.

Run the arithmetic on that. A public poll fielded in the final week, carrying a margin of error of just over two points, missed the final margin by roughly twenty-one. There is no weighting adjustment, no design effect, no rounding convention that gets you from a twenty-point lead to a half-point loss. The poll was not imprecise. It was measuring a different electorate than the one that voted.

Michigan: same error, opposite result

A week earlier, Michigan produced the identical failure in a race the progressive actually won. Emerson College, polling for WLNS and WOODTV on July 26 and 27, had Abdul El-Sayed over Representative Haley Stevens 54 to 39 — a fifteen-point lead, widening to sixteen with leaners. The New York Times noted that after Mallory McMorrow left the race, El-Sayed led by more than ten points in seven of the nine public polls fielded.

El-Sayed won 48.5 percent to 47.5 percent. One point. The winner was correctly identified and the margin was wrong by fourteen. That distinction matters more than it sounds: a poll that names the winner and misses the margin by fourteen points is not a poll that got it right. It is a poll that got lucky about which side of zero it landed on.

The Los Angeles Times, covering the Wisconsin upset, put the two together plainly — polling overestimated progressive performance in both contests. WorldNetDaily was less clinical about it, leading with the observation that the polling was simply wrong and cataloguing the on-air reactions of commentators who had spent weeks treating a Hong nomination as settled. Both are describing the same fact. Only the tone differs.

Minnesota is the control group

Before anyone converts this into a tidy story about progressives underperforming, look at the third race on the same night. In Minnesota, Lieutenant Governor Peggy Flanagan — the progressive — beat centrist Representative Angie Craig 59.0 percent to 39.4 percent. Twenty points, in the other direction. And the June SurveyUSA poll of that race had Craig ahead of Flanagan, 41 to 36. The public polling in Minnesota was wrong about the winner, wrong about the margin, and wrong in the ideological direction opposite to Wisconsin and Michigan. Flanagan is now the nominee to succeed the retiring Tina Smith.

That is the finding that should worry every campaign in the country. If the error ran one way — always toward progressives, or always toward Democrats — it would be a bias, and a bias can be corrected. A pollster can adjust the party-identification target, add a recall-vote weight, tighten the likely-voter screen. What Tuesday produced was not a bias. It was noise of enormous amplitude, pointed in different directions in three adjacent states on the same night, generated by shops using defensible methods.

The mechanism is the one the American Statistical Association warned about long ago and nobody in the business wants to say out loud: when the response rate is low enough, there is no scientifically valid way to infer what the nonrespondents think, and weighting adjustments should not be considered a complete solution. In a general election, high turnout partially rescues the sample — the universe of people who vote is large enough that the willing respondent is not too exotic. In a low-turnout August primary, the person who answers a political survey and the person who drives to a polling place in the middle of summer are two thinly overlapping populations. The 2026 primary polls measured the first group with real precision. The second group picked the nominees.

Polling is bad and it isn't just overpolling progressives.

Minnesota
Poll Average: 🔵Flanagan +2
Actual: 🔵Flanagan +20

Poll Average: 🟡Lindell +7
Actual: 🔴Demuth +6

Wisconsin
Poll Average: 🔴Tiffany +60
Actual: 🔴Tiffany +91

Poll Average: 🔵Hong +24
Actual: 🟡Crowley +0.5…

— America First Insight (@AF_Insight) August 12, 2026

Composition, not preference

The county returns tell the story better than any topline could. Hong won the city of Madison, but she was carrying Dane County — her home county, the University of Wisconsin, the most progressive jurisdiction in the state — by only about 44 to 37 late in the count, far short of the crushing margins the polls implied she was banking. Crowley led the rural counties by six points and the Milwaukee metropolitan area by five, and led narrowly in nearly every county that reported early. Hong did not lose because voters changed their minds about her. She lost because the electorate that showed up was older, more rural, and more institutionally connected than the electorate the surveys had modeled.

This is the argument I have been making with turnout-propensity work for years: a district does not have one partisan lean, it has a different lean at every turnout level. In our legislative modeling, raw registration advantage explained 41.5 percent of the variance in final vote share, while flip scores — the turnout level at which the registration advantage reverses — explained 71.7 percent. Composition beats preference. Wisconsin just paid twenty-one points for the lesson.

One more number from Wisconsin deserves attention, because no poll captures it. About 165,000 votes — nearly 21 percent of the Democratic primary electorate — went to candidates other than the top two, and roughly 68,000 of those went to candidates who had already withdrawn from the race. Crowley’s margin was 3,783. The withdrawn candidates outpolled the winner’s margin by a factor of eighteen. A survey instrument that cannot see a ballot cast for someone who is no longer running is not going to see much else either.

What a wrong number does while it is still believed

Here is the part that goes to the heart of the original argument. The polls in Wisconsin did not merely fail to describe the race. They very nearly determined it.

David Crowley suspended his campaign because his poll numbers were low. He reentered less than a month before election day, after establishment frontrunner Lieutenant Governor Sara Rodriguez dropped out, and then ran an intense late-July effort to consolidate the moderate vote. The candidate who won the nomination had been driven out of his own race by public polling that turned out to be wrong by twenty points. Rodriguez, who quit, still collected more than 32,000 votes from people who had cast ballots before her name came off the board.

On the other side of the ledger, the centrist group Third Way is funding an effort to discredit democratic socialism ahead of 2028, and its spokeswoman had the postmortem ready within hours: Hong “had a whole smorgasbord of positions that were perfectly designed to repel voters in the middle that Democrats have to win.” Perhaps so. But that analysis was equally available on August 10, when the same organization’s allies were reading the same polls and treating the race as lost. Polling did not just report the field. It cleared it, refilled it, and then supplied the moral of the story afterward.

A second instrument has entered the room

They deleted this tweet. pic.twitter.com/BoavMCpSn1

— Christopher Hale (@ChristopherHale) August 12, 2026

The new development this cycle is that polls are no longer the only number the public treats as a forecast. In late July, the Polymarket contract on the Wisconsin Democratic primary was pricing Hong as the favorite to win by fifteen to twenty points, with the next most likely outcomes also Hong victories. By early Wednesday morning, market trackers had Crowley trading near 89 cents on Kalshi — close to the mirror image of the price Hong had carried the night before.

This is not a fringe phenomenon. Traders have moved close to $200 million on the midterms already, by an NBC News analysis reported in the Los Angeles Times. Kalshi has launched a dedicated midterms hub, and its own spokesperson says the platform “has become a leading indicator of where elections are headed.” The problem is what the audience does with the number. A survey commissioned by the Partnership for Large Election Jurisdictions found that 39 percent of likely midterm voters would lose confidence in an election outcome if the official result differed from the prediction-market odds. Three-quarters said the markets create confusion. More than a third believed the odds represented either the current vote count or an official projection from election officials.

Los Angeles County has already lived the consequence. During the June mayoral primary count, market odds helped set an expectation that Republican Spencer Pratt would finish in the top two. He did not, and as the official returns diverged from the market, influencers began alleging fraud. Registrar-Recorder Dean Logan described the position it put his office in: “Not only were we being asked how were the polls wrong … but [people were saying], ‘We saw numbers’ or ‘We saw odds.’” His office is now researching an insider-trading policy for election staff and preparing public messaging for November.

Understand what these markets actually are. They are not an independent check on polling. They are a leveraged bet on polling. The traders who priced Hong near 95 cents were pricing off Marquette and State Navigate. When the polls were wrong by twenty-one points, the market was wrong by roughly the same amount — and it was wrong with money on the table and a veneer of market efficiency that pollsters have never claimed for themselves. The error was not corrected. It was given a second distribution channel.

The count is the next battlefield

Now assemble the pieces as they will present themselves on the night of November 3.

Wisconsin’s primary results were delayed past midnight because roughly 28,000 absentee ballots in Milwaukee were held up by a human error in uploading the data. Nothing improper happened. But hold that fact next to a race decided by 3,783 votes, in a contest where the public had been told for weeks to expect a blowout in the other direction. That is the exact configuration — a confident wrong expectation, a razor-thin result, and a late-night tranche of urban ballots — that manufactures a fraud narrative with no fraud in it.

California enters that configuration with a trust gap already measured. The UC Berkeley Institute of Governmental Studies poll co-sponsored by the Los Angeles Times found that 65 percent of registered California voters are confident local officials will run a secure and fair election, against 31 percent who are not. Split by party, it is 79 percent of Democrats, 62 percent of independents, and 42 percent of Republicans. Split by ideology, it is 82 percent of strongly liberal voters and 38 percent of strongly conservative ones. Berkeley’s Mark DiCamillo called it what it is: “It’s clearly a partisan issue … Strong conservatives and the Republicans are the least confident, and a lot of them are saying [they are] not at all confident.” Election law scholar Rick Hasen supplies the counterweight — “there’s very little evidence of manipulation or of fraud or even of incompetence” — along with the behavioral fact that voters are more confident “when their side wins.”

And California will still be counting. When the Supreme Court heard the Republican National Committee’s challenge to Mississippi’s five-day grace period for late-arriving mail ballots, the six conservative justices sounded ready to require receipt by election day. Justice Samuel Alito framed the stakes bluntly: “We don’t have election day anymore. We have election month or we have election months … Confidence in election outcomes can be seriously undermined if the apparent outcome of the election on the day after the polls close is radically flipped by the acceptance of a big stash of ballots that flip the election.” California counts ballots postmarked by election day that arrive within seven days. In 2024 that was more than 406,000 ballots, about 2.5 percent of the state total.

In the end the Court went the other way. On June 29 it held that the federal election-day statutes do not preempt Mississippi’s law, leaving states free to count ballots postmarked by election day and received afterward. The holding was narrow, and its implications for public confidence are less comforting than the headline suggests. The practical effect here is that the slow California count survives intact. This state will spend days after November 3 adding hundreds of thousands of legally cast ballots to totals that will already have been declared, posted, traded, and litigated in public. The window between the election-night impression and the certified result stays wide open — and that window is exactly where a wrong poll, a mispriced market, and a partisan trust gap compound each other.

Eighty-three days out

The national generic ballot average sits around D+6.7, or D+8.1 once a likely-voter adjustment is applied. Read that number now with Wisconsin in your hand. A state poll fielded in the final week, by a credible operation, in a race with a defined and motivated electorate, missed by twenty-one points. The national average is a slower, broader, better-behaved instrument than any single primary poll — but it is built from the same interviews with the same self-selected respondents, and it is being asked to resolve a House majority that Nate Cohn estimates requires a Democratic popular-vote win of roughly three points just to flip.

So treat the average as a hypothesis rather than a forecast, and watch the four things Tuesday actually validated.

• Turnout composition over topline. Who is modeled into the likely-voter screen, and at what propensity level does the advantage flip? Every August miss was a composition miss, not a persuasion miss.
• Banked votes over projected votes. Early and absentee returns are a census, not a sample. In a state that counts for a week after election day, whoever banks the lead sets the narrative the late count has to fight.
• The residual vote. Twenty-one percent of Wisconsin Democrats voted for someone other than the two contenders, and a third of that went to people who had quit. In November that behavior becomes third-party and undervote share, and no public poll models it honestly.
• The trust differential. Two-thirds of Americans are already concerned that political leaders will pressure election officials over outcomes. A polling miss of Wisconsin’s size, in a swing state, on a Tuesday in November, will not be received as a methodological curiosity.

The paradox, restated

My August note landed on a toss-up, and nothing this week changes that. What changed is the confidence with which anyone should quote a number. Before Tuesday, the case against public polling in 2026 was structural and theoretical — response pools, weighting discretion, the incentive to publish a useful result. After Tuesday it is empirical. In three primaries on one night, in three neighboring states, the public polling missed by roughly twenty-one points, by fourteen points, and by twenty-five points in the opposite ideological direction. Every one of those surveys carried a margin of error near three.

The primaries did not merely disprove the polls. They did something more damaging. They demonstrated that a poll can be off by twenty points and still be reported, aggregated, traded on, and used to drive a serious candidate out of a race — all while wearing a three-point margin of error like a badge of office.

A poll is an argument wearing a lab coat. In August the argument lost twice, won once by accident, and still decided who was allowed to run. Read the crosstabs. Read the turnout tables. Read the top line last, if at all.