Who Is Afraid of the Odds?
Prediction markets, election polls, artificial intelligence, and a problem that will remain long after 2026: who should control the tools citizens use to understand politics?
Prediction markets have become firmly embedded in Brazil’s 2026 presidential election.
On April 14, Flávio Bolsonaro overtook Lula for the first time on Polymarket. At that moment, contracts traded on the platform implied probabilities of roughly 40% for Flávio and 39% for Lula. It was a new signal — and a politically uncomfortable one for the government.
Ten days later, on April 24, Brazil ordered the blocking of 27 prediction-market platforms, including Polymarket and Kalshi. In the same regulatory move, the National Monetary Council restricted derivative contracts tied to political, electoral, sporting, and other non-financial events.
The sequence attracted attention for an obvious reason: a new electoral thermometer began producing an unfavorable result for the incumbent president, and only days later that thermometer was no longer normally accessible in the country.
That does not prove that one caused the other.
Public documents show that the regulatory discussion was already underway. Technical Note SEI No. 2958/2026/MF records that the Secretariat of Prizes and Betting had already identified these markets operating in Brazil, received questions about their legality, and consulted both the Securities and Exchange Commission of Brazil and the Office of the Attorney General for the National Treasury before reaching its final decision. The Secretariat concluded that these contracts displayed characteristics compatible with fixed-odds betting outside the categories authorized under Brazilian law.
There was also an electoral precedent. In September 2024, Brazil’s Superior Electoral Court explicitly stated that betting linked to election results, including wagers placed through online platforms, could constitute electoral violations, abuse of economic power, or illegal vote inducement.
So the story does not need a conspiracy theory to be interesting.
The larger question is this:
what happens when new technologies begin competing with traditional institutions for the right to measure, organize, and interpret political reality?
The 2026 election is only a case study.
The problem will last much longer.
When Opinions Start Costing Money
An election poll essentially asks:
“Who do you intend to vote for?”
A prediction market asks something different:
“What do you believe will happen?”
And the second answer carries financial consequences.
If a contract that pays $1 if a particular candidate wins is trading at $0.58, its price can be read approximately as an implied probability of 58%.
This is not a poll.
There is no representative sample of the electorate.
There are buyers and sellers trying to anticipate a future event.
The central difference is the incentive.
A poll respondent loses nothing for being wrong.
A political commentator can be repeatedly wrong without paying a financial price.
In a market, error costs money.
Guessing is no longer free.
That does not turn prediction markets into oracles. Insufficient liquidity, concentration of capital, herd behavior, manipulation, and privileged information can distort prices.
But they create a different mechanism for aggregating information.
And that mechanism has begun competing with a much older one: the election poll.
From Experiment to Industry
Polymarket was founded in 2020 by Shayne Coplan and grew by turning political, economic, and cultural events into tradable contracts. In 2022, the U.S. Commodity Futures Trading Commission imposed a $1.4 million penalty on the company for offering event-based binary contracts without the required registration.
Kalshi took a different path.
Founded in 2018 by Tarek Mansour and Brazilian entrepreneur Luana Lopes Lara, who met at MIT, it was created around the idea that people should be able to trade directly on future events rather than construct complicated financial instruments to replicate the same risk. Mansour had worked at Goldman Sachs and Citadel; Lopes Lara at Bridgewater and Citadel.
In November 2020, the CFTC designated KalshiEX as a Designated Contract Market, formally integrating it into the U.S. federal derivatives-market framework.
The regulatory contrast between the two countries is revealing.
In the United States, much of the dispute became a question of how these contracts should be classified and regulated.
In Brazil, in 2026, the response was to block the platforms.
When the Thermometers Disagree
On September 20, 2026, Kalshi assigned Flávio Bolsonaro an implied probability of victory of roughly 59%. On Polymarket, he stood at approximately 58%. Lula was near 42% on both platforms.
Those figures did not mean that Flávio had 59% of the vote.
They simply meant that, at that moment, the contracts were trading at prices consistent with a higher expectation of his eventual victory.
Traditional polls showed a much tighter race.
A Quaest poll released on September 21, for example, showed 42% for Flávio and 41% for Lula in a hypothetical runoff, within a two-point margin of error.
So we had two instruments describing the same political process in different ways.
A poll attempts to measure the electorate at that moment.
A market attempts to price the final outcome.
The more interesting question is not simply:
“Which one is right?”
It is:
why do they disagree?
Polls Fail Too
Excessive confidence in election polling also deserves caution.
On the eve of the first round of Brazil’s 2022 presidential election, Datafolha showed 50% of valid votes for Lula and 36% for Jair Bolsonaro. The official result was 48.43% for Lula and 43.20% for Bolsonaro. The Lula estimate was relatively close; Bolsonaro’s result was underestimated by more than seven percentage points.
On the same day, in São Paulo, Datafolha placed Fernando Haddad at 39% of valid votes and Tarcísio de Freitas at 31%.
The final count produced 42.32% for Tarcísio and 35.70% for Haddad. In addition to underestimating Tarcísio by more than eleven points, the poll reversed the order of the two leading candidates.
That does not demonstrate fraud.
It demonstrates something more important:
the published sampling margin of error does not capture every way in which a poll can be wrong.
Nonresponse, weighting, sample distribution, interview methodology, actual turnout, and late shifts in voting behavior can generate errors larger than the margin printed in the footnote.
It is also legitimate to ask who commissioned and financed the survey, how the questionnaire was designed, and how the sample was constructed.
Not because the sponsor proves manipulation.
But because transparency is part of methodology.
The problem begins when an imperfect instrument stops being treated as a measurement and starts being treated as an authority.
Then Artificial Intelligence Entered the Race
In March 2026, Brazil’s Superior Electoral Court established another boundary.
Resolution No. 23.755/2026 prohibited providers of artificial-intelligence systems, even when directly asked by users, from ranking, recommending, suggesting, or prioritizing candidates, parties, campaigns, or coalitions. The rule also prevents such systems from expressing electoral preferences or recommending how someone should vote.
In September, Folha de S.Paulo tested five systems: Claude, DeepSeek, Gemini, ChatGPT, and Grok.
Claude, DeepSeek, and Gemini produced scores or rankings. ChatGPT and Grok refused. When the same questions were repeated 24 hours later, the answers changed and, in some cases, so did the order of the candidates.
The experiment revealed a real weakness of generative models.
But it also revealed a weakness in the question itself.
Asking an AI:
“Who has the best proposals?”
does not define an objective problem.
Best according to what criterion?
Fiscal responsibility?
Economic growth?
Public security?
Healthcare?
Poverty reduction?
Economic freedom?
The environment?
And what weight should each variable receive?
One of the models itself noted during the test that giving equal weight to all categories was an arbitrary choice and that changing those weights would change the ranking.
That point is fundamental.
An AI can answer poorly because the model itself is unstable.
But it can also produce a seemingly precise answer because it was given a methodologically vague question.
A prompt is also a methodology.
The Most Interesting AI Is Not the One That Chooses
The most promising use of artificial intelligence in politics may not be asking:
“Who should I vote for?”
There are far more useful questions.
A structured analysis could compare policy proposals according to predefined criteria: constitutional feasibility, fiscal impact, funding sources, empirical evidence, legal authority to implement them, and the historical record of similar policies.
It could compare current promises with previous decisions.
It could examine successive asset declarations.
It could map declared corporate holdings.
It could reconstruct timelines of public court cases.
It could compare political speeches with documentary records.
And, above all, it could show its sources.
Brazil’s electoral authority makes public data available on candidates, declared assets, social-media accounts, and government programs.
Brazil’s National Council of Justice, through the Public DataJud API, also provides metadata on public court proceedings and procedural developments throughout the country, subject to privacy and secrecy restrictions.
This creates the possibility of a kind of AI-assisted political audit.
But only under one condition: rigor.
Asset growth does not mean illicit enrichment.
Appearance in a court case does not mean guilt.
An investigation is not an indictment.
An indictment is not a conviction.
A first-instance conviction is not the same as a final judgment.
The strength of the tool should lie in preserving these distinctions, not erasing them.
Informing, Analyzing, and Recommending Are Different Things
Comparing two tax proposals is analysis.
Examining declared assets is analysis.
Reconstructing a court case is analysis.
Estimating the cost of a government program is analysis.
Telling the voter which candidate to choose is something else.
This boundary will become one of the most important regulatory questions of the coming years.
Because as AI systems improve, it will make less and less sense to think of them merely as chatbots.
They could become private political research assistants.
Every citizen could, in principle, have something resembling a small research team capable of searching documents, confronting claims, and presenting evidence.
That lowers the cost of investigating power.
And anything that lowers that cost changes politics.
Meanwhile, Ad Hominem Still Works
There is an irony in the debate.
While institutions carefully discuss what prediction markets and AI systems may or may not present to voters, election campaigns continue to rely on a much older and intellectually poorer instrument:
the ad hominem argument.
Instead of showing that a proposal is wrong, attack the person.
Instead of discussing costs and consequences, question character.
Instead of confronting evidence with evidence, create a caricature of the opponent.
That must be distinguished from legitimate scrutiny.
Assets, court records, administrative decisions, conflicts of interest, and public conduct are relevant when someone seeks political power.
The problem begins when personal attack replaces argument.
A broad prohibition on “personal attacks” would be legally dangerous and could suppress legitimate criticism and freedom of expression.
But the paradox remains.
A democracy can design sophisticated rules preventing an AI from generating a ranking while continuing to tolerate campaigns in which insults, insinuations, and slogans receive more attention than proposals, costs, and results.
If the goal is to improve the political education of the electorate, perhaps the relevant question is not only what we should ban.
Perhaps it is the level of quality we should demand from public debate: no ad hominem attacks.
The Real Problem Is the Monopoly on Interpretation
Polls aggregate responses.
Prediction markets aggregate expectations transformed into prices.
Journalists aggregate information and context.
Statistical models aggregate data.
Artificial intelligence can aggregate documents, evidence, and criteria.
None of these instruments has a monopoly on truth.
All of them can fail.
And that is precisely why diversity among them can be valuable.
The democratic challenge should not be to find a perfect machine for political interpretation.
It probably does not exist.
The challenge is to build citizens capable of asking:
Who produced this number?
Using what methodology?
Who financed it?
What was the sample?
How liquid is that market?
What criteria were used?
What was the prompt?
Which documents support the conclusion?
What is fact?
What is inference?
What do we still not know?
The 2026 election will pass.
This problem will not.
Information markets will become more sophisticated.
AI systems will become capable of investigating ever larger volumes of public data.
Governments, courts, and platforms will try to establish boundaries around what these tools may produce.
And citizens will need to learn an uncomfortable lesson:
more information does not automatically mean more truth.
But less information does not produce better citizens either.
The future of democracy may depend less on finding a perfect thermometer than on learning how to compare imperfect ones.
When two thermometers show different temperatures, finding out which one is miscalibrated takes work.
Breaking the thermometer is still easier.
Public Sources
- Ministério da Fazenda — Nota Técnica SEI nº 2958/2026/MF, sobre mercados de previsão e apostas de quota fixa. Serviços e Informações do Brasil
- TSE — Resolução nº 23.744/2024 e normas sobre apostas ligadas às eleições. Justiça Eleitoral
- Reuters — bloqueio brasileiro de 27 plataformas de mercados preditivos, 24 de abril de 2026. Reuters
- CNN Brasil — Polymarket em 14 de abril e Kalshi/Polymarket em 20 de setembro de 2026. CNN Brasil
- CFTC — decisão contra a Polymarket em 2022 e designação da KalshiEX em 2020. CFTC
- Kalshi — história institucional e fundadores. Kalshi
- Datafolha/Folha e TSE — pesquisas e resultados de 2022 usados na comparação histórica. Folha de S.Paulo
- TSE — Resolução nº 23.755/2026, sobre sistemas de IA nas eleições. Justiça Eleitoral
- Folha de S.Paulo — teste com chatbots e candidatos presidenciais, setembro de 2026. Folha de S.Paulo
- TSE — Dados Abertos de candidaturas 2026. Portal de Dados Abertos do TSE
- CNJ — API Pública do DataJud. Conselho Nacional de Justiça
Maurício Veloso Brant Pinheiro, PhD
Professor of Physics, Federal University of Minas Gerais (UFMG)
Founder, Author and Editor, AI-Talks.org
About Us
Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.
Copyright 2026 AI-Talks.org
