With 73 days until the November 3 midterm elections, the country is heading toward the polls with less reliable information about the electorate than at any point in modern memory  and the failures of the past 20 months have come from opposite directions.

In August alone, a fabricated poll in the Los Angeles mayoral race moved money on two prediction markets before anyone checked whether the pollster existed. A week earlier, the entire polling industry missed the Michigan Democratic Senate primary by roughly 12 points. The most respected state pollster in America is still in court over a survey that missed Iowa by 16.

The common assumption is that polls tilt toward Democrats. That was true in 2016, 2020 and 2024. It was decisively false in 2025. The more accurate  and more troubling  conclusion is that the direction of polling error is no longer predictable, which means the number itself now carries less information than the confidence with which it is reported.

Background: Three Cycles of Missing the Same Voters

The American Association for Public Opinion Research (AAPOR), the industry’s professional body, convened a task force after the 2024 election to evaluate 116 general election polls conducted in the final two weeks of the campaign. Its verdict was mixed. State polls were the most accurate since 1944, and national polls performed about average. But for the third consecutive presidential cycle, the polls underestimated Republican vote share.

The task force identified three groups the surveys could not reliably capture: rural, Republican-leaning voters in areas that are hard to reach by traditional methods; Hispanic voters, whose Democratic support was overstated relative to post-election data; and 2024 voters who had sat out 2020, whose share of the electorate was undercounted.

It also found a turnout modeling failure. Counties that backed Donald Trump in 2020 saw turnout surges in 2024, while counties that backed Joe Biden saw declines. Most surveys assumed an electorate distributed roughly like 2020’s. That single assumption explains a meaningful share of the remaining directional error.

In simple terms: the pollsters were not just asking the wrong people, they were guessing wrong about who would bother to show up.

By comparison, the 2020 miss was far larger. AAPOR’s task force on that cycle found the average signed error on the vote margin in the final two weeks ran 3.9 points too favorable to Joe Biden in national polls and 4.3 points too favorable in state polls  the worst performance for national polls in 40 years.

The Mechanics: Why Response Rates Broke the Model

Modern polling rests on an assumption that has quietly stopped holding: that the people who answer are a reasonable stand-in for the people who don’t.

Pew Research Center’s telephone response rate was 36% in 1997 and 6% by 2018. Nate Cohn of The New York Times reported that in 2022, roughly 0.4% of the paper’s dials produced a completed interview.

At those rates, a pollster is no longer measuring the public. It is measuring a small, self-selecting slice of the public and then using statistical weighting to reconstruct everyone else. Weighting works only if the pollster knows what the finished electorate will look like. In a midterm with no presidential race at the top of the ballot, that is a guess.

There is a second, less-discussed failure point: the voter files themselves. Michigan State University political scientist Corwin Smidt, reviewing his state’s primary miss, told WKAR that pollsters were likely modeling the electorate from incomplete or incorrect commercial voter data. He cited his own household as an example  a wife recorded with the wrong party history, a mother missing from the file entirely despite 40 years at the same address.

Case One: Iowa, 2024

On November 2, 2024, the Des Moines Register published its final Iowa Poll, conducted by J. Ann Selzer. It showed Kamala Harris leading Trump by three points in a state Trump had carried by more than eight in 2020. Selzer’s polls carried an A+ rating from Nate Silver’s Silver Bulletin and had correctly called five of the six prior presidential elections in Iowa.

Three days later, Trump won Iowa 55.7% to 42.5%  a margin of more than 13 points, and a miss of roughly 16 points on the margin.

Selzer announced her retirement from election polling two weeks after the election, saying it was unrelated to the poll. In December 2024, Trump sued Selzer, Selzer & Co., the Register and parent company Gannett under the Iowa Consumer Fraud Act, alleging the poll constituted “brazen election interference” rather than an error. The defendants deny the claims and argue the suit is barred by the First Amendment.

The litigation has been procedurally contested. Trump dropped the federal case in June 2025 and refiled in Polk County District Court hours later  one day before Iowa’s new anti-SLAPP statute took effect. In November 2025, the Eighth Circuit let that move stand, keeping the case in Iowa state court, where it remains pending.

The legal question is whether an inaccurate poll can be treated as consumer fraud. That is a novel theory, and most election-law scholars have been skeptical of it. The journalistic question is separate and unresolved: a single outlier from a trusted pollster drove three days of national coverage, and no correction mechanism existed until the votes were counted.

Case Two: The 2026 Primaries

The 2026 primary calendar has produced misses larger than anything in 2024  and they have run against the party’s left flank, not against Republicans.

Michigan, August 4. Abdul El-Sayed defeated Rep. Haley Stevens in the Democratic Senate primary by about one point, fewer than 15,000 votes out of more than 1.5 million cast. The RealClearPolitics average had El-Sayed ahead by 10.2 points. Analyst G. Elliott Morris put the pre-election average at roughly 13 points, making the industry’s collective error about 12.

Adam Carlson of Zenith Research, whose final survey for a pro-El-Sayed super PAC showed a 53–34 lead, attributed the error to differential non-response and to mistaking enthusiasm for turnout  the gap between how easy it is to take an online poll and how much effort it takes to vote. An Emerson College survey for 6 News and WOOD TV8 had El-Sayed up 54–39.

Wisconsin, August 11. State Rep. Francesca Hong, endorsed by the Democratic Socialists of America, led the gubernatorial primary polling by roughly 20 points. She lost to Milwaukee County Executive David Crowley by less than one. CNBC reported that prediction markets were wrong in the same direction and by a similar magnitude, which undercuts the argument that betting markets correct for polling error. They largely import it.

Neither result is a partisan error in the conventional sense. Both are failures to model who actually turns out in a low-participation primary  and in both, the candidate running to the left was overstated.

Case Three: Los Angeles, August 2026

The Los Angeles case is different in kind. There was no methodology to critique.

An outfit calling itself Median Strategies published a survey showing Mayor Karen Bass leading Councilmember Nithya Raman by roughly 12 points ahead of the November 3 runoff, including a claim that Bass had absorbed the vast majority of voters who backed eliminated Republican candidate Spencer Pratt in June. The Los Angeles Times reported on August 17 that the numbers were fabricated. Median Strategies replaced its site with a notice calling the results a “short-term social experiment.” The Guardian reported the operation was run by a 21-year-old recent college graduate.

The poll was picked up by television and radio stations, at least one national outlet, and Bass’s own campaign, which promoted it in a since-deleted post. A campaign spokesperson said staff reviewed the accompanying memo and judged it legitimate in part because other outlets had already reported it.

The market response is the part worth pausing on. The Associated Press tracked thousands of Polymarket trades from roughly 20 accounts within six minutes of the campaign’s post. On Kalshi, a “yes” contract on Bass rose from 63 to 65 cents within 15 minutes.

AAPOR condemned the fabrication as a serious breach of ethics. Mark Baldassare, survey director at the Public Policy Institute of California, told CalMatters he had never encountered a wholly invented survey in decades of work and called it a wake-up call. David Dayen of The American Prospect wrote that the episode proved a pump-and-dump scheme transfers cleanly from penny stocks to election markets.

The Counter-Case: 2025

Any argument that polling systematically favors Democrats has to account for last November.

In New Jersey, Gov. Mikie Sherrill beat Jack Ciattarelli by double digits more than the final polling average, according to Sabato’s Crystal Ball. In Virginia, Abigail Spanberger won by 15.4 points, exceeding the pre-election average by several points, with Democrats outrunning the polls in all three statewide executive races. The analysis attributed much of the New Jersey error to underestimating Democratic margins among minority voters. Nate Silver made the same point more bluntly: sometimes polls underestimate Democrats.

New York City that year cut both ways. Three-way polls overstated Zohran Mamdani’s eventual winning margin; two-way polls understated it. And in the June 2025 Democratic primary, a Marist survey a month out had Andrew Cuomo leading Mamdani 37% to 18%, with a simulated ranked-choice run ending 60–40 for Cuomo. Mamdani won the final round 56.4% to 43.6%.

The through-line from Iowa 2024 to Wisconsin 2026 is not partisan tilt. It is that pollsters cannot currently model who votes when the composition of the electorate is in flux.

Impact: What This Means for November

The generic congressional ballot illustrates the problem. As of mid-August 2026, the major averages do not agree on the size of the Democratic advantage:

Source	Democratic margin
Morning Consult (registered voters)	+4
ActiVote (likely voters, Aug. 9)	+5
Decision Desk HQ average	+6 (47–41)
Silver Bulletin average	+6.7 raw; +8.1 likely-voter adjusted
Emerson College (single poll, July)	+11

Sources: Morning Consult, ActiVote, Decision Desk HQ, Silver Bulletin, Emerson College via The Hill.

A seven-point spread between reputable sources is not noise around a known number. It is disagreement about what the electorate is. Under the new Texas and Missouri congressional maps, analysts have noted that Republicans could lose the national House popular vote by roughly a point and still hold the chamber  meaning a four-point average and an eight-point average describe two different outcomes, not two versions of the same one.

That uncertainty propagates. Donor allocation, super PAC spending, party committee triage of which districts to defend, candidate recruitment and media narratives all run on polling averages. When the averages are wrong by 12 points, as they were in Michigan, every downstream decision made on them was also wrong.

Analysis

Three structural changes deserve more attention than they get.

Polls now move money directly. Prediction markets have made survey results a tradable input. That creates an economic incentive to publish favorable numbers and, as Los Angeles showed, to invent them. Baldassare’s objection was not only that the fake poll deceived voters  it was that opinion research now drives financial decisions it was never built to support.

Verification failed at every layer. The Median Strategies numbers passed through newsrooms, broadcast outlets and a sitting mayor’s communications shop without anyone confirming the firm existed. The campaign’s stated reason for trusting it was that other outlets had reported it  a circular standard that will fail again.

The error is not stable, so “correcting” for it is guesswork. Pollsters who adjusted after 2020 to avoid understating Republicans were positioned to understate Democrats in 2025. Anyone applying a mechanical correction to 2026 numbers based on 2024 is fighting the last war.

There is a partial silver lining in the calendar. As Baldassare noted, the Los Angeles fabrication surfaced in August, when public attention is low  not in late October with ballots in hand.

Conclusion

Polls remain the best instrument available for understanding what voters care about. They are considerably weaker as forecasts, and weaker still in primaries and off-year contests where turnout is unpredictable.

The practical standard for readers over the next 10 weeks is not to dismiss polling but to interrogate it: who conducted it, who paid for it, how many respondents, likely voters or registered voters, what electorate the weighting assumes, and whether the pollster discloses methodology at all. A single survey  even from a firm with a decades-long record, as Iowa demonstrated  is not evidence of a trend.

The 2026 midterms will be decided by turnout in an electorate that no one has successfully modeled in three consecutive attempts. Anyone claiming otherwise is selling confidence the data does not support.

Key Takeaways

• AAPOR found 2024 polls were more accurate than 2016 or 2020, but underestimated Republican support for a third straight cycle, driven partly by a turnout model that assumed a 2020-style electorate.
• The final 2024 Des Moines Register/Selzer Iowa Poll showed Harris +3; Trump won by more than 13. The resulting consumer-fraud lawsuit is pending in Iowa state court.
• The 2026 primaries produced larger misses than 2024: roughly 12 points in Michigan’s Democratic Senate primary and roughly 20 in Wisconsin’s gubernatorial primary  both overstating the candidate to the left.
• A fabricated Los Angeles poll fooled news outlets and the Bass campaign in August 2026 and moved prices on Polymarket and Kalshi within minutes.
• Polling error is not reliably partisan. Polls understated Democrats in New Jersey and Virginia in 2025 and understated Mamdani in the New York City primary.
• Telephone response rates fell from 36% in 1997 to 6% by 2018; roughly 0.4% of New York Times dials produced an interview in 2022.
• Generic ballot averages currently disagree by about seven points, a spread wide enough to describe entirely different outcomes for House control.

Related Coverage

Blakeman’s Rise in the Polls vs. Hochul: What the Siena Numbers Actually Show  NexfinityNews (verify slug against the CMS before publishing).

Sources

• American Association for Public Opinion Research, Task Force on 2024 Pre-Election Polling (October 29, 2025); Executive Summary.
• AAPOR, statement on fabricated polling data (August 2026).
• Karlyn Bowman and Ruy Teixeira, “Election Polls: Their Past, Present, and Future,” Forbes/AEI (November 10, 2025).
• CalMatters, “A fake poll fooled the LA mayor’s campaign,” (August 20, 2026).
• Associated Press, reporting on Median Strategies and prediction-market activity (August 2026).
• Los Angeles Times, initial report on the fabricated Bass–Raman survey (August 17, 2026).
• The Guardian, reporting on the operator of Median Strategies (August 20, 2026).
• CNBC, “Massive polling misses in Wisconsin, Michigan primaries” (August 13, 2026).
• WLNS, “Michigan’s primary polling miss, explained by a pollster who got it wrong” (August 2026).
• Washington Examiner, “What went wrong with the polls in the Michigan Senate race?” (August 5, 2026).
• WKAR Public Media, interview with Corwin Smidt, Michigan State University (August 18, 2026).
• Sabato’s Crystal Ball, “Polling Accuracy in the 2025 Elections” (December 11, 2025); “Digging Deeper into Virginia, New Jersey Returns” (December 15, 2025).
• Silver Bulletin, “Actually, sometimes polls underestimate Democrats” (January 2026).
• Iowa Public Radio and NBC News, coverage of Trump v. Selzer (December 2024–July 2025); Washington Examiner on the Eighth Circuit ruling (October 2025).
• Marist College Institute for Public Opinion, New York City Democratic primary survey (May 2025).
• Pew Research Center telephone survey response-rate trends, 1997–2018, via NBC News.
• Decision Desk HQ, Silver Bulletin, Morning Consult and ActiVote generic ballot averages (August 2026).