Your Chatbot Now Quotes the Betting Odds

The Odds Moved Into the Chatbots

X named Polymarket its official prediction market in June 2025, with Grok writing explanations of the odds.¹ Kalshi added Grok on July 24, 2025.²

Google came next, in November 2025. "We're also adding support for prediction markets data from Kalshi and Polymarket, so you can ask questions about future market events and harness the wisdom of the crowds," Google said of its AI-upgraded Google Finance.³

OpenAI's first reported deal arrived in July 2026. ChatGPT search showed Kalshi's World Cup odds, such as France at 59% against Spain.⁴ Bets still happen on Kalshi's own app.

Kalshi uses AI too. It reportedly built an agent on Anthropic's Claude, called Harrison, to test contract wording before markets open.⁶

Wall Street pays for the pipe. Intercontinental Exchange, owner of the New York Stock Exchange, agreed in October 2025 to invest up to $2 billion in Polymarket.⁷ ICE also signed on to distribute Polymarket's data worldwide.

Dow Jones signed an exclusive Polymarket data deal in January 2026. "We're making prediction markets data accessible to our users, because it's a rapidly growing source of real-time insight into collective beliefs about future events," said Almar Latour, the chief executive of Dow Jones.⁸

Latour picked the word beliefs. A belief is a guess somebody holds with confidence.

Cable news signed on in December 2025. CNN's segment The Odds has featured Kalshi at least 115 times, Public Notice counted.¹⁷ Some CNBC stories carry this disclosure: "CNBC and Kalshi have a commercial relationship that includes customer acquisition and a minority investment." In at least 22 cases, CNBC left it out, Public Notice found.¹⁷

How Big the Bet Got

Growth came in fits, then all at once. The 2024 election lifted combined monthly volume to $4.74 billion in November 2024.⁵ It sagged to $1.72 billion by February 2025 and sat at $2.02 billion that July. December 2025 reached $11.98 billion. In May 2026 it hit $25.66 billion, and sports trading during the World Cup doubled it by July.

August cooled things. Volume dipped to about $47 billion, the first monthly drop in a year, by Pew's count.⁵ An earlier count from The Block put it at $45.33 billion.⁹

Add up Pew's months, and the pair traded about $259 billion in contracts from January through August. That is five times all of 2025.

Those dollars need a warning label. Each contract pays $1 if it wins, and Pew counts every contract at that $1.¹⁰ Buy a long shot for 5 cents, and the tally still books a full dollar. Volume overstates the cash at risk.

Investors paid up anyway. On Aug. 31, 2026, Bloomberg reported that Polymarket will raise $1 billion at a $21 billion valuation, in a round led by 1789 Capital.¹¹ Kalshi, last valued at $22 billion, is in talks to raise money at $40 billion.¹²

Trained, Tested or Just Shown?

Here is the part the hype skips. "AI trained on prediction markets" can mean three different things.

Trained means a model learned from the data. A 2025 study trained a 14-billion-parameter model on 10,000 settled Polymarket questions plus 100,000 machine-generated ones.¹³ It beat OpenAI's o1 at forecasting. Polymarket's own prices beat it, though: a Brier score of 0.151 for the market against 0.190 for the model, where lower is better.

Why train a model on bets at all? A settled market makes a tidy answer key, since every question settles one way or the other. In that study, the model read news articles while the market price stayed hidden from it. It was scored on how each question settled.

Tested means a benchmark grades a model on market questions. ForecastBench, from the Forecasting Research Institute, draws its market questions from Kalshi, Metaculus and Polymarket.³⁷ Its Oct. 13 leaderboard mixes those with questions built from public data.¹⁴ Overall, a Google DeepMind entry scored 69.0%. The superforecasters' median forecast scored 68.6%, and the public's median forecast 63.4%. Saying 50% on every question earns 50%.

That lead sits inside the margin of error. On market questions alone, the superforecasters still lead, 74.7% to 73.7%, with the public at 68.2%.¹⁴ The institute's baseline board runs models with zero tools or added context. There, the best model scores 61.9%, below the public's 62.5%.¹⁴ In January, the institute projected a tie between models and superforecasters in November 2026, with a range running to November 2027.¹⁵

Shown means a chatbot repeats the odds to you. Grok and Google Finance live there. ChatGPT showed Kalshi's odds during the World Cup.

The AILately.com Desk found zero public statements from OpenAI, Google, Anthropic, xAI or Meta about training a flagship model on Kalshi or Polymarket data. Grading is another story. xAI entered Grok in ForecastBench itself.¹⁵

When the Crowd Is Wrong, or Cheating

A market price comes from whoever trades. Columbia researchers estimated that wash trading, where a trader buys from themselves, made up 25% of Polymarket's activity over three years.¹⁶ The study has yet to face peer review.

"Wash trading doesn't add liquidity or information to the market, so it would seem valuable to distinguish authentic from inauthentic volume," said Yash Kanoria, a Columbia Business School professor and co-author.¹⁶ He named information. Fake trades inflate the volume and may move the price that anyone reads, a chatbot included.

Thin markets swing on small money. A Kalshi market on Greenland rose from 30% to 39% on $24,000 of "yes" trades, Public Notice found. CNN presented that market's 43% odds on air.¹⁷

Then come the insiders. A Polymarket trader placed bets totaling $32,000 that Venezuela's Nicolás Maduro would fall and won more than $400,000.¹⁸ In April 2026, prosecutors charged an Army special forces soldier with trading on the secret raid.

"Prediction markets are not a haven for using misappropriated confidential or classified information for personal gain," said Jay Clayton, the U.S. attorney for the Southern District of New York.¹⁸

A Google software engineer faces fraud charges after allegedly making $1.2 million on bets about Google's own Year in Search list.¹⁹ OpenAI fired an employee in February 2026 for using confidential information on prediction markets.²⁰

AI is a betting market too. Polymarket's market on the best AI model at the end of 2025 drew $36.3 million and settled on LMArena's public leaderboard.²¹

How Far Into Gambling Does This Go?

Far. From July 2024 to early May 2026, sports made up 80% of Kalshi's volume and 39% of Polymarket's, Pew found.¹⁰ During the World Cup, in June and July, sports trading topped $58 billion on Kalshi alone.⁵

Americans bet about $40 billion at legal sportsbooks in the first quarter of 2026. Prediction-market sports trading was about on par, Pew found.⁵

Courts disagree about what this is. The Third Circuit sided with Kalshi in April, finding federal law likely governs its sports contracts.²³ Later, the Ninth and Sixth circuits sided with the states.²⁴²⁵ Arizona filed 20 criminal counts against Kalshi in March.²⁴

Washington has stepped back. The Commodity Futures Trading Commission has about one-eighth the staff of the SEC, NPR reported.³⁵ Trump officials dropped several federal investigations into the industry, which now runs more than 200,000 active markets.³⁶ Donald Trump Jr. advises both Kalshi and Polymarket.¹⁸

"Kalshi's position would have effectively federalized sports betting regulation nationwide and swept aside decades of state and tribal oversight built to protect consumers from the real harms of unregulated gambling," said Arizona Attorney General Kris Mayes.²⁴

Problem-gambling experts use the same word. "Regardless of how prediction markets are currently legally defined, NCPG believes it is functionally gambling and can expose consumers to many of the same risks and harms associated with traditional gambling," said Derek Longmeier, board president of the National Council on Problem Gambling.²⁶ Kalshi gave the council $2 million.²⁶

Does anyone win? Pew tracked 11,989 Polymarket accounts for six weeks this year.²⁷ The average trader spent about $600 and lost less than $2. Add Pew's loss brackets, and 56% of accounts lost money.

Robots Are on the Board

Machines show up as the thing being bet on. Polymarket's market on a driverless Tesla robotaxi service before July 2025 drew $11.9 million and resolved against Tesla.²⁸ This month, traders priced Tesla releasing its Optimus robot by the end of 2026 at 7.25 cents on the dollar.³⁴

A search by The AILately.com Desk turned up zero public cases of a robotics company training on market odds. Those odds put a price on what traders expect Tesla's robot to do by the end of 2026.³⁴

The Case For and the Case Against

Start with the strongest case for the data. Ethereum co-founder Vitalik Buterin saw AI as the missing piece. "AI changes that equation completely, and means that we could potentially get reasonably high-quality info elicited even on markets with $10 of volume," he wrote in 2024.²⁹

Federal Reserve staff agree on the value. "Our results suggest that Kalshi markets provide a high-frequency, continuously updated, distributionally rich benchmark that is valuable to both researchers and policymakers," wrote Anthony M. Diercks and two colleagues.³⁰ The Fed notes that staff papers speak for their authors.

The case against rests on what fills a price: wash trades, leaked secrets and sports bettors. Kanoria's 25% figure and Clayton's indictment both describe prices built on something besides honest forecasting.

Both things can be true. Markets can see the future partly because some traders already know the answer, and some can shape it. Kalshi fined three federal candidates in April for betting on their own races.²² A confident price can also miss.

Questions Readers Ask

Can AI help with prediction markets? For research, yes. The Polymarket-trained model above earned a simulated return above 10%.¹³ It still trailed the market.

Which AI is best at prediction? An unnamed Google DeepMind entry leads ForecastBench at 69.0%. Grok 4.20 scores 67.7%, OpenAI's gpt-5.5 67.4% and Anthropic's Claude Opus 5 66.6%.¹⁴

Can ChatGPT predict the stock market? Partly, and briefly. In one study, GPT-4's reads of news headlines matched the market's first reaction on about 90% of portfolio-days.³² Its authors found the edge shrinks as more traders use the tools.

What does training frontier AI cost? Epoch AI found the cost of the biggest training runs grows 2.4 times a year. It expects the largest to pass $1 billion by 2027.³³

Can you make money on prediction markets? Some people do. In Pew's sample, 44% of accounts made money, and 7% made more than $1,000.²⁷

Is AI 100% right? Far from it. Vectara measures how often models invent facts while summarizing a document. Rates run from 1.8% to 24.2%, with GPT-5.5 at 9.3%.³¹

What to Watch

Supporters need cleaner markets, with fewer wash trades and fewer insiders. Critics need proof that bad odds reach real models and real wallets. New Jersey has asked the Supreme Court to review its Kalshi case.²⁵ Kalshi's talks at $40 billion will show how much Wall Street believes.¹²

A price on Kalshi or Polymarket is a guess with money behind it. People guess wrong, and machines that learn from those guesses will too.

Sources

1. Liz Napolitano, "Polymarket to Serve as Official Prediction Market of Elon Musk's X," Decrypt, June 6, 2025.
2. Chris Eudaily, "Musk's Grok AI is now on the Kalshi, Polymarket betting apps. What could go wrong?" CNBC, July 25, 2025.
3. Logan Hitchcock, "Google Finance Integrates Polymarket, Kalshi Prediction Market Data," Decrypt, Nov. 6, 2025.
4. Helen Partz, "OpenAI quietly adds Kalshi World Cup odds to ChatGPT: Report," Cointelegraph, July 14, 2026.
5. Kaitlyn Radde, "Prediction markets' trading volume doubled between May and July, largely driven by sports," Pew Research Center, Sept. 23, 2026.
6. Todd Shriber, "Kalshi Using Claude AI to 'Referee' Prediction Markets," Casino.org, June 15, 2026.
7. Intercontinental Exchange, "ICE Announces Strategic Investment In Polymarket," via Mondo Visione, Oct. 7, 2025.
8. Brad Senkiw, "Polymarket Increases Media Exposure with New Dow Jones Partnership," Covers, Jan. 7, 2026.
9. Timmy Shen, "Kalshi and Polymarket's combined volume falls 15% in August, first monthly decline in a year," The Block, Sept. 2, 2026.
10. Kaitlyn Radde, "Trading volume on prediction markets has soared in recent months," Pew Research Center, May 27, 2026.
11. Amanda Gordon, "Polymarket Funding Led by 1789 Values Firm at $21 Billion," Bloomberg, Aug. 31, 2026.
12. Felix Ng, "Kalshi in advanced talks to raise new funding at $40B valuation: Reuters," Cointelegraph, Sept. 30, 2026.
13. Benjamin Turtel, Danny Franklin, Kris Skotheim, Luke Hewitt and Philipp Schoenegger, "Outcome-based Reinforcement Learning to Predict the Future," arXiv, Dec. 1, 2025.
14. Forecasting Research Institute, "Leaderboards," ForecastBench, Oct. 13, 2026.
15. Houtan Bastani, Simas Kučinskas and Matt Reynolds, "LLMs Are Closing the Gap on Human Superforecasters," Forecasting Research Institute, Jan. 29, 2026.
16. Lydia Beyoud, "Polymarket volume inflated by 'artificial' activity, study finds," Fortune, Nov. 7, 2025.
17. Aaron Rupar and Judd Legum, "How Kalshi infects the news," Public Notice, July 6, 2026.
18. Bobby Allyn, "U.S. soldier charged with using classified information to bet on Maduro's removal," NPR, April 23, 2026.
19. Aaron Katersky and Katherine Faulders, "Google employee charged with fraud after D4vd search bet won $1M on Polymarket," ABC News, May 28, 2026.
20. Julie Bort, "OpenAI fires employee for using confidential info on prediction markets," TechCrunch, Feb. 27, 2026.
21. Polymarket, "Which company has best AI model end of 2025?" Polymarket, Dec. 31, 2025.
22. Eva Roytburg, "Kalshi and Polymarket are racing to ban insider trading. The economist who built the theory behind prediction markets says it's the whole point," Fortune, April 26, 2026.
23. Johnny P. ElHachem, "Federal Appeals Court: CFTC Jurisdiction Over Sports Event Contracts Likely Exclusive," Holland & Knight, April 7, 2026.
24. Public Gaming Research Institute, "Arizona: Ninth Circuit ruling backs state authority over Kalshi sports contracts, but election dispute remains," Public Gaming Research Institute, Sept. 2, 2026.
25. Cheryle Shepstone, "Sixth Circuit Rules Against Kalshi, Lets Ohio and Tennessee Enforce Sports Betting Laws," DeFi Rate, Sept. 25, 2026.
26. Sean Chaffin, "National Problem Gambling Group Calls Prediction Markets 'Gambling'," Card Player, Sept. 28, 2026.
27. Kaitlyn Radde, "What we know about the typical Polymarket user," Pew Research Center, July 22, 2026.
28. Polymarket, "Will Tesla launch a driverless Robotaxi service before July?" Polymarket, July 1, 2025.
29. Vitalik Buterin, "From prediction markets to info finance," Vitalik Buterin's website, Nov. 9, 2024.
30. Anthony M. Diercks, Jared Dean Katz and Jonathan H. Wright, "Kalshi and the Rise of Macro Markets," Federal Reserve Board, February 2026.
31. Vectara, "Hallucination Leaderboard," Vectara, Sept. 22, 2026.
32. Alejandro Lopez-Lira and Yuehua Tang, "Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models," arXiv, Oct. 28, 2025.
33. Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej and David Owen, "How much does it cost to train frontier AI models?" Epoch AI, June 3, 2024.
34. Polymarket, "Will Tesla release Optimus by...?" Polymarket, Oct. 13, 2026.
35. Bobby Allyn, "A $400,000 profit on Maduro's capture raises insider trading questions on Polymarket," NPR, Jan. 5, 2026.
36. Bobby Allyn, "Kalshi reveals insider trading case against editor for MrBeast," NPR, Feb. 25, 2026.
37. Forecasting Research Institute, "About," ForecastBench, Oct. 13, 2026.