Prediction markets and AI are creating a new loophole in election polling
A new study warns that platforms such as Polymarket can influence campaigns while avoiding rules governing traditional election surveys.
New tools that appear to the public as polls and can similarly influence election campaigns, yet are not subject to the transparency rules and restrictions that apply to election polls, are creating a loophole that should be closed before the upcoming elections. That is the conclusion of a study by Dr. Tehilla Shwartz Altshuler, Dr. Guy Lurie and Dr. Rachel Aridor-Hershkovitz of the Israel Democracy Institute.
The study examines two new “poll alternatives”: prediction markets such as Polymarket and AI-based simulations. Although neither is a public opinion poll, both can be presented to the public using metrics that resemble polling results, including probabilities of winning, projected parliamentary seats and levels of support.
According to the researchers, this raises two concerns: methodological errors and inaccuracies, and the emergence of alternatives that can carry the professional authority associated with polls without being subject to the same disclosure requirements and publication restrictions.
2 View gallery
From right: Prime Minister Benjamin Netanyahu and Yashar Chairman Gadi Eisenkot
(Shalev Shalom, Shaul Golan)
The study emphasizes that Polymarket is not a poll. An election poll seeks to measure preferences, who people intend to vote for, while a prediction market prices expectations about future outcomes. Its participants do not constitute a representative sample, and the “one person, one vote” principle does not apply. Participants who invest more money can have a greater influence on the market price.
The researchers found that even relatively small sums can shift the “win probabilities” displayed by Polymarket, and that the platform is not consistently more accurate than polls or models based on them. Its performance is particularly poor in markets with low trading volumes.
The researchers also warn about the potential for media manipulation. A price shift lasting only a few hours can be captured in a screenshot and turned into a headline claiming that a candidate has “surged to a 70% chance of winning.” This can create a feedback loop: the market reacts to news and campaigns, and the resulting market price is then reported by the media as if it were fresh, independent information about the state of the race.
According to the researchers, presenting prediction markets as a reliable source of election information risks blurring the distinction between speculative betting, in which participants risk money based on their assessment of a future event, and opinion polls, which measure voter sentiment.
The researchers reach a similar conclusion about AI-based “surveys.” They say there is currently no basis for treating them as reliable substitutes for human surveys for the purpose of statistical inference. Studies have found that while synthetic averages may resemble human responses, they can exhibit lower variance and different correlations between variables. Minor changes to a prompt or model version can also alter the results.
The researchers therefore propose treating Polymarket data and AI simulations as survey substitutes under election laws whenever they are presented as measurements or quantitative forecasts of voting patterns, parliamentary seats or winning probabilities.
The proposal would impose two key restrictions. New data could not be published from the Friday preceding the opening of polls until the polls close, consistent with the rules governing traditional election surveys. During the rest of the election period, publishers would have to provide transparency and disclosure about how the data was generated.
Under the proposal, anyone publishing data based on Polymarket would have to disclose trading volume, liquidity, the number of participants, market concentration and anomalous transactions, while making clear that the data does not represent a representative sample.
For AI-based surveys, publishers would have to disclose, among other things, the model used, data sources, prompts, number of runs and weightings, while clarifying that the “respondents” are not human beings.
The study also cites findings from the Anti-Corruption Data Collective, a nonprofit research organization. The research, reported by Reuters on September 9, examined more than 11,000 markets related to congressional elections on Polymarket and Kalshi. It found that in 94% of the markets examined, a single trade of less than $1,000 could shift the price by an amount equivalent to a 10-percentage-point change in probability. In hundreds of cases on Polymarket, the new price persisted for at least a full day.
Both platforms dispute the implication that such movements undermine the usefulness of prediction markets. They argue that a distorted price creates an arbitrage opportunity: if a market price diverges from the probability other traders consider plausible, those traders can bet against the distortion and profit when the price corrects itself. In their view, this creates a market incentive to restore prices to more accurate levels. Reuters reported similar arguments from both platforms.
Another study cited in the paper found that 1% of wallets accounted for approximately 68% of trading volume in the congressional markets examined, while just 10 wallets accounted for about 17%. The researchers argue that this means a relatively small number of participants can exert significant influence over the “odds of victory” presented to the public.
The researchers also cite an example of a potential forecasting failure. Last August, Polymarket and Kalshi gave a candidate in the Wisconsin primaries a 95% to 96% chance of winning, according to the study, but she ultimately lost.
The Central Elections Committee declined to comment on the Israel Democracy Institute's study.
