The Machine Midterms: AI Agents and the 2026 Election Odds

The First Election With a Scoreboard Nobody Elected

On November 3, tens of millions of Americans will vote in the midterm elections. Long before the first polls close, a different kind of result will already be scrolling across cable chyrons and social feeds: a live probability, updated by the second, for control of the House and the Senate. That number will not come from a pollster. It will come from a pair of prediction markets, Kalshi and Polymarket, where real money prices the odds of almost every competitive race in the country.

Commentators have taken to calling this the first prediction-market election, and the label is earned. The odds are everywhere now, cited by anchors and campaign strategists as if they were returns themselves. The deeper change, though, is not reach. It is authorship. The scoreboard that pundits treat as the aggregated wisdom of a crowd is, increasingly, the output of software trading against other software.

The stakes have grown to match the attention. Combined open interest on midterm contracts across the two venues reached roughly $250 million by October 18, with about $210 million of it sitting on Kalshi, per data cited by crypto.news. That is not play money, and Washington has noticed. The Commodity Futures Trading Commission, which regulates these contracts, says it is watching both platforms for manipulation heading into the vote. Three forces now converge on the same screen: AI agents that trade the races, institutions that have bet billions on the data, and courts that cannot agree whether any of it is legal.

This piece is about that collision. Not the horse race, but the machine that prices it: who really moves the odds, how those odds can be gamed, who decides when a bet has been won, and what could break on election night when a country reads its own future off a market that mostly trades itself.

Who Is Actually Trading the Midterms

Start with the traders, because they are not who most readers imagine. The retail punter betting a hunch on his party still exists, but he is no longer the marginal price-setter. That role has shifted to automated agents: software that reads headlines, prices probabilities, and fires orders without a human in the loop. On Polymarket, more than 30% of active wallets are run by such agents, and 14 of the 20 most profitable wallets are bots, according to CoinDesk’s reporting on the platform. Polystrat, an agent launched there in February 2026 by the firm Valory, placed more than 4,200 trades in its first month and booked up to 376% on a single position.

David Minarsch, Valory’s chief executive, describes Polystrat as “an autonomous AI agent that trades on Polymarket 24/7 on behalf of its human user.” He is candid about the ceiling: simply prompting an off-the-shelf model with a market, he has said, usually produces outcomes no better than a coin flip. The edge comes from custom workflows that pull data, weigh sources, and keep positions current around the clock. The result, in his telling, is blunt: agents tend to do better than humans. The platform’s own numbers back him up, with roughly 37% of these agents running a positive profit and loss against something like 7% to 13% of human accounts.

Those agents do not live in browser tabs. They run from programmable smart-account wallets that can hold funds, sign transactions, and operate under spending limits set by their owners, the same account architecture explored in our guide to what you actually own with smart-account wallets. That is what lets a bot trade a Senate contract at 3 a.m. with no one awake to approve it.

The midterm book is also strikingly concentrated. On Polymarket, 24,930 wallets held a position in the election contracts, with a median stake of just $14, yet the top 100 wallets controlled 56% of all position value, per the figures cited by crypto.news. The two venues split along familiar lines, with Kalshi taking about 84% of midterm open interest and Polymarket the rest. A thin layer of large, often automated accounts sets the price that a long tail of tiny human bets rides along with. When people say the market thinks a race is a coin flip, this is the market they mean.

How the Odds Became the Story

Prediction markets did not earn their place on the chyron by accident. They earned it in November 2024, when they called a presidential race the polls had framed as a coin toss. Polymarket priced a Donald Trump victory near 95% hours before any network projected it, and more than $3.5 billion flowed through its 2024 presidential market, as NBC News documented at the time. For a generation of traders and reporters, that was the moment the markets stopped being a curiosity and started being a forecast.

The legal path ran alongside it. Kalshi spent 2024 fighting the CFTC for the right to list election contracts at all, and a federal court cleared the way that autumn, making regulated election betting a fixture rather than a gray-area experiment. By 2026, the question was no longer whether markets would price the midterms, but how loudly. Bernstein analysts now project the sector will grow from about $51 billion in volume in 2025 to roughly $240 billion in 2026 and $1 trillion by 2030, an argument laid out in CNBC’s coverage of the firm’s research.

That success created a feedback loop. The more outlets cite the odds, the more the odds matter; the more they matter, the more capital flows in to move them. A probability that was once a quiet aggregation of private beliefs is now a public object that campaigns, donors, and algorithms all react to in real time. The danger in that loop is the subject of the rest of this article, because a number that influential is a number worth attacking.

The Ghost of 2024: One Whale, One Narrative

Before trusting the crowd, it helps to remember how small the crowd can be. The headline figure of the 2024 cycle was not a bot but a man. A French trader known only as Theo wagered around $28 million that Trump would win, spread across roughly 11 Polymarket accounts, and stood to collect as much as $85 million, NBC News reported. In late October 2024, a single large order was enough to swing the market’s implied Trump probability sharply in the moment it landed.

Theo turned out to be right, and he had done real homework, even commissioning his own polls on a thesis that public surveys were undercounting Trump voters. But the episode exposed a structural truth that carries straight into 2026: a market probability can be the conviction of one well-funded actor rather than the consensus of thousands. France’s gambling regulator opened an inquiry into Polymarket in the aftermath. The platform had, by one reading, forecast the election; by another, it had broadcast a single whale’s bet to the world as if it were the voice of the crowd.

Now layer the agents on top. Where Theo was one human with a thesis, 2026 adds dozens of automated accounts that can take correlated positions in milliseconds, reinforce one another’s moves, and chase the same signal into the same thin order book. The concentration problem has not gone away. It has been automated, and it has been handed a faster trigger finger.

Mention Markets and the Insider Problem

The purest manipulation risk is not in the big partisan contracts but in the small, oddly specific ones: will a candidate say a certain word in a debate, will an official attend an event, will a figure be named in a speech. These so-called mention markets settle on human conduct, which means the people closest to that conduct hold information nobody else can price. In September 2026, the CFTC issued a staff advisory warning that contracts tied to an individual’s words, attendance, or behavior can, in its phrasing, create heightened manipulation risks, and urging platforms to build contract-specific safeguards, as summarized by crypto.news.

The warning was not theoretical. Two cases this year put names to the problem. Gabriel Perez, a White House technical adviser who operated the presidential teleprompter, used his advance view of prepared remarks to trade Kalshi contracts on what the president would say. The CFTC ordered him to return $107,539 in profits, pay a $65,000 penalty, and accept a three-year ban, a roughly $172,000 reckoning detailed by NBC News. Weeks earlier, former Representative George Santos drew the first lifetime ban Kalshi has ever issued, alongside CFTC penalties, over positions tied to whether he would attend the 2026 State of the Union, as Axios reported.

Here is the paradox that defines these markets. Every trade settles on a public blockchain, so Polymarket is arguably the most surveilled venue in finance; anyone can watch a suspicious wallet load up before news breaks. Yet transparency catches wrongdoing after the fact; it does not prevent the trade. Detection is a forensic exercise, the kind of chain analysis that firms profiled in our look at crypto’s audit toolchain have turned into a discipline. Prevention would require knowing who holds an edge before they act, which no amount of on-chain visibility delivers. For an election, where the insiders are campaign staff, speechwriters, and officials with early access to returns, that gap is not a footnote. It is the whole exposure.

The Manipulation Playbook for an Election Market

Insider trading is only one page of the playbook. A market that moves headlines can be attacked from several directions at once, and the 2026 cycle has previewed most of them.

• The fake signal. In August, a 21-year-old published fabricated polls, later calling them a short-term social experiment, and watched them ripple through news coverage and market odds after a candidate amplified one showing phantom momentum, a sequence NPR traced in detail. A market that prices the news can be moved by manufacturing the news.
• Volume painting. Trading that sits on both sides of a contract can manufacture the appearance of conviction. Analysts have flagged clusters of oddly uniform order sizes on both venues as a tell that the activity is not what it looks like.
• Coordinated agent flow. Dozens of bots keyed to the same data feed can pile into a thin swing-state contract together, pushing the implied odds well past what the underlying evidence supports before human traders react.
• Resolution gaming. If you cannot move the price, you can attack the settlement, betting that an ambiguous outcome will be resolved your way and leaning on the process to get there.

Agents amplify every item on that list. They act faster than any human desk, they correlate because they often read the same sources and run similar logic, and they feed on exactly the kind of machine-readable signal a fake poll produces. The surveillance side is automating too, with on-chain monitors that score wallets for suspicious patterns, but the contest is asymmetric: the attacker needs one opening, while the defender has to watch every market at once.

Can a Bot Forecast an Election Better Than the Market?

If agents dominate the trading, a fair question follows: are they actually better forecasters, or just faster ones? The most rigorous answer so far comes from Prophet Arena, a benchmark from the University of Chicago that grades frontier language models against real Kalshi and Polymarket questions, using the market price as a contamination-proof ground truth. Its published results are humbling for the hype. The best models now match the market on calibration, meaning their stated probabilities are about as well-tuned as the market’s own, but not one of them reaches break-even when it actually bets.

The reason is temperament. The models are, in the authors’ words, consistently conservative; even when a market prices something as near-certain, the model hesitates and refuses to commit to the extreme. Calibration is not conviction, and conviction is where the money is. The same models stumble on exactly the inputs an election demands: recalling time-stamped facts, weighing contradictory sources, and updating fast as returns come in. The raw intelligence behind these agents keeps improving, powered by the kind of distributed model-serving described in our report on decentralized inference, but intelligence and edge are not the same thing.

That gap matters for the midterms. A market full of well-calibrated but timid forecasters can be accurate on average and still get pushed around at the margin by a single aggressive actor, human or machine, who is willing to take the extreme position the models will not. The crowd of bots is cautious. The whale is not. It does not take much to guess which one sets the headline number in a thin market at midnight.

Who Settles an Election Bet

A prediction market is only as trustworthy as its referee, and on election night the referee is the whole game. The two venues settle disputes in opposite ways. Kalshi resolves centrally: it checks the result against a named source agency filed with regulators, the Associated Press for most races, and routes any dispute to its own Outcome Review Committee, whose decision is binding with no independent appeal. Polymarket outsources the job to UMA, a decentralized optimistic oracle where anyone can propose an outcome by posting a bond, and token holders vote only if someone disputes it, a process its documentation lays out in full.

Both models have a timing problem that elections make acute. A race can be called by networks hours or days before it is certified, and a contract’s fine print decides which moment counts. In past cycles, Polymarket has publicly overridden its own oracle when a vote produced a result it deemed wrong, which quietly turns a supposedly automated referee into an advisory one. Stanford political economist Andrew Hall, writing for a16z, argues the fix is to specify the judge in advance. At contract creation, he proposes, “the market maker specifies not just the resolution criteria in natural language, but the exact LLM and the exact prompt that will be used to determine the outcome,” so that, as he puts it, “the entire resolution mechanism is visible and auditable before anyone places a bet.”

That is a clean idea for a clean question. A contested House seat headed for a recount is not a clean question, and no oracle, human or machine, can resolve a result that does not yet exist. Resolution is the one risk a trader cannot hedge, and it is the risk most likely to surface after the polls close.

When the Odds Move the Voters

So far the worry has run one way: can bad actors move the market? Election officials are worried about the reverse, that the market will move the election. When live odds become a national scoreboard, they stop being a neutral mirror and start shaping behavior, from turnout to trust.

Reporting by Votebeat captured the fear plainly. Dean Logan, the chief election official in Los Angeles County, described a local contest where, in his words, “early election returns differed from market expectations, and that resulted in suspicion and questioning of normal ballot processing.” When the count and the odds disagree, he warned, it is the count that gets doubted, because “elections administration right now is very much affected by perception and misinformation.” His office saw exactly that during a Los Angeles mayoral primary in June 2026.

Others see a more direct threat. Jim Allen, an elections director in Pennsylvania, recounted being pitched on letting officials place small bets on turnout and refusing flatly: “And we said, no, this is all bad.” Wisconsin invoked a 19th-century statute barring voters from wagering on races they vote in, which Kalshi called active voter suppression. The common thread is that a financial market now sits on top of a civic process, and the incentives of the two do not line up. A market rewards a confident number. A democracy needs an accurate count, even when it is slow and boring.

The Money Behind the Scoreboard

None of this would matter if the venues were still hobby projects. They are not. Intercontinental Exchange, the owner of the New York Stock Exchange, has committed about $2 billion to Polymarket across an initial stake in October 2025 and a further $600 million in March 2026, a deal The Block reported turns ICE into the global distributor of Polymarket’s event data. Crucially, ICE’s thesis is not betting. It is the data: a real-time, money-weighted probability feed on politics and the economy that it can sell to the same institutions that already buy its market data.

That reframes the midterms. For the professionals, the election odds are not a casino line; they are a sentiment instrument, a hedge, and a signal to wire into other trades, including the crypto and macro positions laid out in our year-end market outlook. The appetite is real and the growth is steep.

One footnote keeps the hype honest. The infrastructure tokens underneath these markets remain tiny. UMA, the oracle that settles Polymarket, carries a market value in the low tens of millions of dollars, and Olas, the network behind many trading agents, is smaller still. The platforms are worth billions; the crypto rails they run on are worth a rounding error of that. It is a useful reminder that the value here is in the exchange and its data, not in the governance tokens that trade on their names.

The Legal Cloud Over Election Night

For all the money, the single most important fact about these markets is that no one can agree they are legal everywhere. Prediction markets live under the CFTC, not the SEC; the event contracts are federal commodities instruments, and the SEC only enters where crypto tokens settle them. But a growing list of states say election and sports contracts are gambling, their turf, and both venues already block users in roughly nine states.

The fight is now in open court. On October 16, Polymarket sued Ohio in federal court, arguing that the state’s licensing demands discriminate against interstate commerce and are preempted by the Commodity Exchange Act, a filing Crypto Briefing covered as it landed on the day Ohio’s compliance deadline expired. Above it all sits the Supreme Court: New Jersey’s petition against Kalshi is docketed as No. 26-299, and Kalshi’s response is not due until November 9, as Sportsbook Review noted, which is to say after the midterms are over. The country will vote, and the market will price the vote, before the highest court even decides whether to hear the case.

Underneath the venue fight is a classification fight. In October the CFTC moved to treat many event contracts as swaps, which would pull them fully under federal derivatives law, even as three federal appeals courts split on whether sports and election contracts are swaps at all. The patchwork that results is more than a lawyer’s problem, as our guide to following the regulatory chokepoints explains. It is the exact condition that shapes who can trade, and how.

Why One Rulebook Is the Agents’ Precondition

Here is where the legal story and the machine story become the same story. A human trader can read a map of which states are off-limits and simply avoid them. An autonomous agent cannot, not cleanly. Software does not carry a driver’s license; it routes around geofences, it runs from servers that could be anywhere, and it scales by doing the identical thing in every jurisdiction at once. A 50-state patchwork is a nightmare for a bot in a way it never is for a person.

That is why federalization, a single national rulebook under the CFTC, is not a side issue for the agent economy. It is the precondition for it. One clear federal standard is what would let automated trading scale without tripping over a different gambling statute every few hundred miles. The institutions pouring billions into these venues are, in effect, betting that Washington wins the turf war, because only a federal market is a market their machines can trade at size.

The irony is that the referee is shrinking just as the game grows. NPR reported this month that the CFTC is stretched thin even as prediction markets and crypto balloon under its remit. The agency says it is monitoring the platforms for manipulation ahead of the midterms, but watching a book that trades hundreds of millions of dollars, much of it machine-driven and around the clock, is a tall order for an under-resourced regulator. The machines are already here. The rulebook is still being written.

What Could Break on Election Night

Put the pieces together and a few failure modes become easy to picture once the polls close.

• The divergence. A swing state’s early returns contradict the market’s confident odds. Instead of concluding the market was wrong, a slice of the public concludes the count is rigged. This is the scenario election officials fear most, and it needs no manipulation at all to happen.
• The thin-book swing. In a lightly traded contract at 2 a.m., a cluster of correlated agents or a single whale moves the implied odds 20 points on a rumor. Screenshots travel faster than the correction.
• The resolution standoff. A race goes to a recount, and a seven-figure contract cannot settle for days. On Polymarket that means a drawn-out oracle dispute; on Kalshi, a closed-door committee decision that losing traders will not trust.
• The late insider case. A wallet that loaded up before a results leak gets flagged, and the story becomes not who won, but whether the market was ever clean.

None of these are exotic. Each has a rehearsal somewhere in the 2026 record already, from the Los Angeles primary to the fake-poll stunt to the Perez and Santos cases. The midterms simply run them all at once, under maximum attention, on a book that mostly trades itself.

The Scoreboard Nobody Elected

The honest verdict on prediction markets is neither the booster’s nor the scold’s. The markets are genuinely informative: they aggregate real money and real conviction, and in 2024 they read the country better than the polls did. They are also genuinely gameable, concentrated, machine-dominated, and dependent on a settlement layer that strains under contested outcomes. Both things are true at once, and the midterms will show both.

The AI-agent lens is what makes 2026 different from 2024. Agents now set much of the price, agents forecast the races (and lose money proving they are too timid to beat the market they match), agents hunt for manipulation, and agents may yet move the very odds that voters and anchors treat as truth. Transparency helps catch the cheating after it happens; it does not stop it. Resolution and reflexivity, not raw accuracy, are the risks that will define election night.

So watch the number on November 3, but watch it the way you would watch any instrument that mostly trades itself: as a signal, not a verdict. The crowd it claims to speak for is smaller, faster, and more automated than the word crowd suggests. The country has, almost without deciding to, installed a new scoreboard for its own elections. Nobody voted for it, and on election night the whole machine gets its first national stress test.

Frequently Asked Questions

Are prediction markets legal in the US for the 2026 midterms?

Yes, at the federal level. Kalshi and Polymarket operate as CFTC-regulated event-contract venues, and a 2024 court ruling cleared the way for regulated election contracts. But several states treat them as gambling, so both platforms block users in roughly nine states, and the Supreme Court has been asked to decide whether states can apply their own gambling laws. Kalshi’s response in that case is not due until November 9, after the election.

Do AI agents really control prediction-market prices?

They dominate a large and growing share. On Polymarket, more than 30% of active wallets are autonomous agents and 14 of the 20 most profitable wallets are bots. Humans and institutions still provide plenty of volume, but automated accounts increasingly set the marginal price, especially in fast-moving, short-dated contracts where speed matters most.

Are prediction markets more accurate than polls?

They were in 2024, when Polymarket priced the presidential result well ahead of the networks. But accuracy is not the same as being manipulation-proof. A single large trader can move a headline probability, liquidity in minor contracts is thin, and benchmarks show that even top AI models match the market’s calibration without beating it. Treat the odds as one informative signal, not a settled forecast.

Can you be punished for insider trading on a prediction market?

Yes. The CFTC treats event contracts as its domain and has brought cases. In 2026 it penalized a White House teleprompter operator who traded on advance speech drafts and sanctioned former Representative George Santos over a State of the Union attendance market, which also drew the first lifetime ban Kalshi has issued. Contracts tied to a person’s words or conduct are under particular scrutiny.

What happens if a race is too close to call?

Settlement can stall. Kalshi resolves against a named source such as the Associated Press and sends disputes to an internal committee, while Polymarket relies on the UMA oracle, where a contested outcome can take days to resolve through a token-holder vote. A recount or a late call can leave a contract unsettled well after election night, which is the one risk a trader cannot hedge.

Marcus Okafor covers the collision of artificial intelligence and crypto markets for HOGE Wire.