On November 3, 2026, Americans voted in the midterm elections for the House of Representatives, a third of the Senate and many governorships. As of the evening of November 4 (Vienna time), reports summarised on X say Democrats won control of both the House and the Senate. Counting is still going on in some races, and official certification takes weeks.
Before every election, an industry of forecasters makes predictions: pollsters, statistical models, betting and prediction markets, pundits, and, increasingly, AI tools. This post takes a first look at who got it right and what role AI played.
Important note on sources: This article is written less than 24 hours after polls closed. Most of the information comes from news summaries on X, which are generated from posts and can contain errors; X itself marks them with “Grok can make mistakes, verify its outputs.” We have not checked results against official state election data. Treat all numbers here as preliminary.
The result in brief (preliminary)
According to X news summaries from November 4, 2026:
- Democrats won majorities in both chambers of Congress.
- They reportedly flipped seats in states considered Republican strongholds, including Texas, Kansas, South Carolina and Iowa.
- Before polls closed, independent models had projected around a 54–46 Democratic Senate majority; on election night, one summary cited model projections of 54 Senate and 249 House seats for Democrats.
Final seat counts were not available in the sources we read. We will not give exact numbers until official results are in.
Contestant 1: The pollsters
Polls were the classic forecasting tool. Before the election, polling averages showed Democrats ahead on the “generic ballot” (which party voters prefer for Congress). One X summary cited an average of 8.7 points in Silver Bulletin’s average at the end of October.
AtlasIntel tops first ranking: According to an X summary updated on November 4, the Brazilian polling firm AtlasIntel ranked first among 32 pollsters for its national generic ballot forecast, which projected a 13.8-point Democratic lead. The summary says AtlasIntel correctly called all 10 Senate races and six governor races it surveyed. CEO Andrei Roman highlighted Iowa, Kansas and Maine. AtlasIntel was also praised for its 2020 and 2024 results. We could not find who produced the ranking of 32 pollsters or what method it used, so this claim is unverified.
The “recall bias” debate: On October 31, forecaster Chaz Nuttycombe argued that polls might overstate Democrats by about two points because some 2024 Trump voters do not report their past vote correctly, which distorts the weighting. If the preliminary results hold, this warning did not materialise; the error appears to have gone in the other direction. A full analysis will need final numbers.
Contestant 2: Statistical models
Forecast models such as Nate Silver’s combine polls, past results and economic data into probabilities. According to X summaries from November 2, professional models favoured Democrats to win both chambers in most simulations. Historian Allan Lichtman, known for his “13 Keys” system for presidential races (which missed in 2024), predicted over 90% odds for a Democratic House win and 67% for the Senate, likely 51–49. If Democrats end up near 54 seats, his direction was right but he underestimated the margin.
Contestant 3: Prediction markets
Prediction markets let people trade contracts that pay out if an event happens. Prices are read as probabilities. Kalshi and Polymarket are the biggest platforms in the US election space.
- On election night, Polymarket gave Democrats around 74% for Senate control and was quoted at 93% for the House, according to X summaries.
- According to a summary updated on November 4, Kalshi markets “nailed 479 of 487 races,” a 98% hit rate, with Polymarket reaching hundreds of millions in daily volume and Kalshi more than $2.5 billion in total volume.
Caution: A 98% hit rate sounds impressive, but most of the 487 races were not competitive. A model that simply predicts the incumbent party in safe districts already gets most of them right. The meaningful test is the close races and how well-calibrated the probabilities were (did events priced at 70% happen about 70% of the time?). We found no independent analysis of this yet. The source of the 479/487 figure is also not clear from the summary.
Contestant 4: Pundits on X
Many confident predictions on X were far off. According to X summaries:
- Some Republican-leaning accounts predicted 55 Republican Senate seats and a 10-seat House majority.
- Commentator Will Chamberlain forecast a D+4 environment, with Republicans keeping 53 Senate seats and Democrats winning the House 220–215.
- Pollster John McLaughlin predicted a Republican Senate hold at 53 seats.
- Mark Halperin, citing private sources in both parties, spoke of a “blue tsunami” late on November 3.
If preliminary results hold, partisan predictions on both sides performed worse than markets and models, a pattern seen in earlier elections too.
Where does AI come in?
The user question behind this article was: were AI-based predictions correct? The honest answer today is: we do not know yet, and there is little hard data. Here is what we found:
1. AI is built into the forecasting tools
On October 30, a new prediction market called Aries launched, focused only on elections and economic outcomes, with regulated yes/no contracts cleared through ForecastEx. According to an X summary, it integrates AI so that users can query live market data via tools like Claude or ChatGPT. Aries showed Democrats at 72% to win the Senate before the election. Here AI is a user interface, not the forecaster itself.
Pollsters, too, increasingly use machine learning for weighting and modelling. AtlasIntel describes its methods as proprietary; we do not know to what extent they rely on AI.
2. LLMs as forecasters: no systematic test found
Researchers have tested whether large language models can predict events, for example in forecasting tournaments, with mixed results: models have become better but generally have not clearly beaten the best human forecasters or markets. For the 2026 midterms we found no published, systematic comparison of chatbot predictions with the results. Individual users posted chatbot forecasts on X, but cherry-picking is easy: the correct ones get shared, the wrong ones are forgotten. We therefore do not make claims about how well “AI” predicted this election.
3. AI as a political issue
AI also played a role as a topic. According to an X news summary from November 4, an executive order (named as EO 14434, dated September 29, 2026) had directed federal agencies to replace “artificial intelligence” and “AI” with “Super Intelligence” and “SI” in documents, without changing definitions. The summary says the rebrand was reversed after Democrats won Congress, sparking jokes and a rush on .si domains. We could not verify the executive order or the reversal in an official source for this article, so this is unverified and reported here only as X discussion.
More broadly, AI costs and jobs were part of the political background. X discussions in the days before and after the vote focused on rising electricity and hardware prices linked to AI data centres, and on warnings about entry-level jobs. Whether these issues influenced votes would require exit poll data that we did not review.
4. Markets react, including tech
Nasdaq futures fell by about 1% on election night as Democrats took the lead, according to X summaries, on expectations of higher taxes and stricter regulation for tech. For AI companies, a Democratic Congress could mean new debates on AI safety rules, data centre energy use and antitrust. What actually happens depends on legislation still to come.
What we can learn so far
- Aggregated signals beat loud individuals. Markets, averages and models pointed in the right direction; many confident pundits did not.
- Hit rates are misleading. “98% correct” says little without looking at close races and calibration.
- Polling error is not fixed. Before the vote, some warned of a pro-Democratic error; the preliminary results point the other way.
- AI is mostly an interface so far. In 2026, AI helped people access and explain forecasts, but we found no evidence that a chatbot-based forecast was systematically tested.
- Wait for final numbers. Rankings published within 24 hours are based on incomplete counts.
What we could not verify
- Final seat counts in the House and Senate.
- The ranking of 32 pollsters and the methodology behind AtlasIntel’s claimed accuracy.
- Kalshi’s “479 of 487 races” figure.
- The “Super Intelligence” executive order and its reversal.
- Any systematic test of AI chatbot predictions for these midterms.
Conclusion
The first winners of the 2026 midterm forecasting contest, according to reports on X, are the aggregated tools: prediction markets, polling averages and some pollsters like AtlasIntel. AI was everywhere around the election, from chatbot interfaces on betting platforms to political debates about AI itself, but as a forecaster it still lacks a proper scorecard. That is the article worth writing once final results and independent analyses are available.
Sources
All X links below are news summaries generated from posts on X; X labels them as possibly inaccurate.
- X: Prediction Markets Accurately Forecasted 2026 Midterm Winners
- X: AtlasIntel Tops Accuracy Rankings for 2026 Midterm Polls
- X: Democrats Lead in Key 2026 Midterm Races per Prediction Markets
- X: Prediction Markets Favor Democrats in Tight 2026 Senate Midterms
- X: Aries Launches Election Prediction Markets
- X: Election Forecaster Warns of Recall Bias Boosting Democrats
- X: Allan Lichtman Predicts Democrats Will Take House and Senate
- X: Conservative X Forecasts Predict Narrow GOP Senate Hold
- X: Pollster McLaughlin Predicts GOP Senate Hold
- X: Halperin Predicts Democratic Wave
- X: Stock Futures Drop as Democrats Lead
- X: Federal ‘Super Intelligence’ Rebrand Ends After Democratic Election Win
- Kalshi · Polymarket
- Wikimedia Commons: South Dakota “I voted” sticker 2026 (image)
