Industries / Government, education and community / Campaigns and polling
Polling methods
Polling is asking a sample of people how they will vote, or how they just voted, and adjusting their answers so that they stand for everyone who votes; on election night the same statistics project the result from early returns.
6
Milestones achieved, from 1952 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1952
A computer projects the count
On election night CBS News teletyped early returns to Remington Rand's factory in Philadelphia, where a UNIVAC I computer compared them with the vote expected in each place from past elections and projected the rest. Early in the evening it pointed to a landslide for Dwight Eisenhower, but its programmers held the forecast back because pre-election polls had shown a close race. Eisenhower won by a landslide, and only hours later did CBS tell viewers the machine had been right.
LessonThe forecast was held back because it disagreed with the polls, and the polls were the ones that were wrong.
Credit
CBS News; Remington Rand
1967
Voters asked at the door
For CBS News's coverage of Kentucky's race for governor, Warren Mitofsky and Murray Edelman interviewed voters as they left their polling places. The Roper Center calls it the first major exit poll for a news network; the first national exit poll followed in 1972, and other networks later adopted the method.
Credit
Warren Mitofsky and Murray Edelman, CBS News
2000
A state called, then retracted
On election night the US networks, all fed by Voter News Service (VNS), a consortium of five networks and the Associated Press that ran the exit poll and counted sample precincts, called Florida for Al Gore, retracted it, called it for George W. Bush, and retracted that too. A report for CNN by Joan Konner, James Risser and Ben Wattenberg found "serious flaws" in VNS's exit polling, models and technology, and told CNN to stop using exit polls to call states.
LessonAn exit poll "will never be as accurate as a properly conducted actual vote count," so close races wait for counted votes.
Credit
Konner, Risser and Wattenberg, report for CNN, 2001
2018
Voter surveys replace exit polls
Because more Americans were voting early or by mail, the Associated Press replaced the in-person exit poll with AP VoteCast, built with NORC at the University of Chicago and used by Fox News and other outlets. It combined a probability-based, state-by-state survey of registered voters with a large opt-in online survey, and its first edition interviewed 113,677 voters and 21,599 nonvoters. In November 2025 AP and six television networks merged their efforts into The Voter Poll, run by SSRS, which blends a probability survey of people confirmed to have voted with in-person exit interviews.
LessonWhen voting moved away from the polling place, the exit poll had to follow it.
Credit
AP and NORC; SSRS for The Voter Poll
2020
Australia's polls fail together
Every national poll published during Australia's 2019 federal campaign had Labor ahead on the two-party-preferred vote, the count after smaller parties' preferences are shared between the two main sides; the Coalition won with 51.5%. An inquiry led by the Association of Market and Social Research Organisations (AMSRO), reporting on 11 November 2020, called it a "polling failure": the samples leaned toward "more politically engaged and better educated voters," were not adjusted for it, and all erred the same way. It ruled out a late swing and "shy" conservative voters as causes.
LessonPolls that use different methods can share one bias, so agreement among them is not proof.
Credit
AMSRO inquiry
2025
Smaller errors, same lean
A task force of the American Association for Public Opinion Research (AAPOR) found that 611 US polls from the last two weeks of the 2024 campaign missed the two-party margin by 3.3 percentage points on average, down from 5.3 in 2020 and 5.2 in 2016; state presidential polls were the most accurate since 1944. For the third presidential cycle running they underestimated Republican vote shares relative to Democrats', by 2.7 points on the margin, and the task force found no evidence of "herding," pollsters nudging results toward each other's.
LessonWeighting on how people voted last time narrowed the spread between polls, but "no single methodological recipe guarantees higher accuracy.".
Credit
AAPOR Task Force on 2024 Pre-Election Polling
Watch for
Markets as election forecasts
Prediction markets, exchanges where people trade contracts that pay out on an election result, say their favorites won almost every race in the US elections of 3 November 2026. Kalshi counted 479 wins in 487 settled races, and a CoinDesk review of the platforms' own figures put the share at about 98%; each platform chose its own races, so the figures cannot be compared with each other or with polls.
Achieved whenan independent group scores market prices and polling averages on the same races at the same points before the vote, with a calibration measure such as the Brier score, and publishes the data.
Claimed by
Kalshi and Polymarket, November 2026
Synthetic electorates
Aaru, a US startup, builds AI agents meant to stand in for voters and surveys them in place of people. Fortune reported that Aaru's simulation of New York City's mayoral primary came within 2,000 votes of the final count; the poll aggregator FiftyPlusOne said in March 2026 that it will leave out any "poll" with AI-generated respondents.
Achieved whensynthetic forecasts lodged publicly before several elections are scored by an outside party against the results and against real polls of the same races.
Claimed by
Aaru, June 2026
Open
01
Poll errors by method, worldwide
An open table of final pre-election polls from national elections around the world, 2018 to 2026, each with the official result and the poll's method coded: how people were interviewed (phone, online, face to face or mixed), how they were recruited (at random or opt-in), and what the answers were weighted on. Then a published analysis of which choices go with smaller errors outside the US.
The briefCredit · unclaimed02
Seal synthetic forecasts first
A public contest in which makers of synthetic polls, plus an open-model baseline, lodge sealed forecasts for coming elections before voting ends, built only from aggregate public statistics. An outside team opens them after the polls close and scores them against the results and against the final average of real polls.
The briefCredit · unclaimed
Open 01 · the brief
Poll errors by method, worldwide
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open table of final pre-election polls from national elections around the world, 2018 to 2026, each with the official result and the poll's method coded: how people were interviewed (phone, online, face to face or mixed), how they were recruited (at random or opt-in), and what the answers were weighted on. Then a published analysis of which choices go with smaller errors outside the US.
Why it is still open
The largest cross-national record, Will Jennings and Christopher Wlezien's 30,000-plus polls from 351 elections in 45 countries, ends in 2017, and its public data on Harvard Dataverse were last revised in September 2020. AAPOR's 2025 test of methods against error covered US polls only, and found most choices "showed little relationship to error size". Open poll archives on GitHub cover single countries, such as US Senate and German Bundestag polls up to 2021.
Built with what exists
national opinion-poll lists as a starting index, each poll checked against the pollster's own release; official results from election commissions; an AI model to pull method details out of pollsters' PDF reports, checked by hand on a sample; Python or R.
First step
Code every poll from the final week of five recent national elections in countries with different polling traditions, and publish the codebook.
Done looks like
At least 3,000 final-month polls from at least 100 elections in 40 countries, each method field traced to a pollster document; two coders agreeing on at least 90% of a 200-poll sample; average error by mode and recruitment reported with 95% confidence intervals; data and code public.
Open 02 · the brief
Seal synthetic forecasts first
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public contest in which makers of synthetic polls, plus an open-model baseline, lodge sealed forecasts for coming elections before voting ends, built only from aggregate public statistics. An outside team opens them after the polls close and scores them against the results and against the final average of real polls.
Why it is still open
Most published checks of synthetic samples are made after the answer is known. Aaru reruns clients' old surveys "blind against the known result", and an open validation set built from the 2025 Canadian Election Study scores models against answers already collected. The forecasts made in advance come one team and one election at a time: Shapeng Jiang, Lijia Wei and Chen Zhang of Wuhan University posted a forecast of the 2024 US presidential election from ChatGPT-4o "virtual polls" on 3 November 2024, two days before the vote, with no other entrant scored beside it.
Built with what exists
open-weight and commercial AI models through their APIs; census tables and past official results; SHA-256 hashes of each forecast posted to a public repository before the vote, with the forecast itself released only after polls close so that it cannot sway anyone; Python for scoring.
First step
Publish the scoring rules, the hashing procedure and the list of elections due in the first half of 2027 before any forecast is made.
Done looks like
Sealed forecasts for at least 10 national or state elections, each entrant's average error on vote shares set beside the real-poll average's error for the same races, all published within a month of the last vote.