# Can AI predict elections? What the evidence shows in 2026

> AI can forecast races, and one model called 2024 a week early. But AI "synthetic polls" miss real ones by 12 points. What works, what fails and why.

Oct 1, 2026 · AI forecasting · Sikt Intelligence · https://www.siktintelligence.com/blog/can-ai-predict-elections

**Short answer: partly.** AI can put useful probabilities on elections, and the best AI forecasters now hold their own on real-world questions. But the most hyped shortcut, using AI to fake polls, does badly: in a September 2026 test by Pew Research Center, AI "respondents" differed from real polls by 12 percentage points on average. The honest version of "predicting an election" is a probability, made before election day and checked across many races.

## Three ways AI tries to predict elections

### 1. Simulating voters

The idea: give a language model the profile of a real voter (age, region, education, past views) and ask how that person would vote. Do it thousands of times and you have a "silicon sample". In a well-known 2023 study, researchers found that GPT-3, conditioned on backstories from real US survey participants, could "accurately emulate response distributions from a wide variety of human subgroups", a property they called algorithmic fidelity ([Argyle et al., *Political Analysis*](https://arxiv.org/abs/2209.06899)).

### 2. Forecasting with research

The second approach treats an election like any other forecasting question. An AI system reads polls, news, history and [base rates](/blog/base-rates-forecasting), then gives a probability. This is how [AI superforecasters](/blog/what-is-an-ai-superforecaster) work, and on general real-world questions the best of them are now statistically tied with human superforecasters on one major benchmark (see [can AI predict the future?](/blog/can-ai-predict-the-future)).

### 3. Combining the signals

Elections come with unusually rich data: polling averages, forecast models, early-vote numbers and [prediction markets](/blog/what-is-a-prediction-market). A good forecaster, human or AI, weighs all of them rather than trusting one. We compare the two most common in [prediction markets vs. polls](/blog/prediction-markets-vs-polls).

## What happened in 2024

One study offers a rare clean test. Researchers used GPT-4o to simulate voters based on data from the American National Election Studies, finished their forecast on or before October 30, 2024, and posted it on November 3, two days before election day. Depending on the method, it predicted a Trump win with roughly 286 to 309 electoral votes ([Jiang, Wei and Zhang, 2024](https://arxiv.org/abs/2411.01582)). Trump won with 312.

That is a genuine forecast, made before the result was known. But it is one election. A race that looks close is not far from a coin flip, and a coin flip calls the winner half the time. To know whether an AI forecasts elections well, you need many races, scored with a proper measure such as the [Brier score](/blog/brier-score-explained) and checked for [calibration](/blog/forecast-calibration).

The big AI companies took the risks seriously. In March 2024, Google restricted its Gemini chatbot from answering many election-related questions, including requests to predict winners, "out of an abundance of caution" ([9to5Google](https://9to5google.com/2024/03/12/google-gemini-election-questions/)).

## Where AI falls short: fake polls

Simulating voters works far less well than its fans hope. In September 2026, Pew Research Center compared AI-generated "synthetic" survey answers with its own human polls across three survey waves from early 2026. The synthetic results differed from the human results by an average of 12 percentage points, with the largest errors for Republicans (16.1 points), Black adults (15.1) and adults without a college degree (13.6). Pew's verdict: "the AI survey respondents did not match the views of their human counterparts especially well" ([Pew Research Center](https://pewresearch.org/data-labs/2026/09/30/how-well-synthetic-samples-replicate-public-opinion)).

A 12-point error is larger than the margin in almost every competitive race. Synthetic polls are not a substitute for asking real people.

## Why elections are a good test for AI forecasting

- **They resolve clearly.** There is a winner on a known date, so every forecast can be scored.
- **The evidence is rich.** Polls, fundamentals, fundraising and history all inform the answer.
- **There is a benchmark.** Prediction markets price the same races, so an AI forecast can be compared with the market in real time.
- **The outcome does not adapt to your forecast.** Unlike stock prices, which change as soon as a signal is widely used ([can AI predict the stock market?](/blog/can-ai-predict-the-stock-market)), a good forecast of a Senate race does not make the race harder to forecast.

The hard part is volume. A national election comes every two years, so building a track record takes many races, not one headline.

## How to judge an AI election prediction

1. **Was it made before election day,** with a timestamp? Anything published afterwards proves nothing.
2. **Is it a probability,** not just a pick? "Candidate A, 62%" can be scored; "Candidate A wins" cannot be judged fairly.
3. **Does it cover many races?** One correct call is luck until proven otherwise.
4. **Is it compared with a baseline,** such as polling averages and prediction markets?
5. **Is it scored after the result,** with misses included?

## Where Sikt Intelligence fits: the Midterm Bench

Sikt Intelligence is putting its AI superforecaster to exactly this test. In the **Sikt Midterm Bench**, it is forecasting the key 2026 Senate races ahead of election day, independently of the markets. As November 3 approaches, we will publish its forecasts on our [2026 midterms page](/midterms), next to Kalshi and Polymarket, and score every one against the results.

Until then, you can follow the live market odds for every close race, from [Maine](/odds/maine-senate-race-odds) to [Texas](/odds/texas-senate-race-odds), and for [control of the House and Senate](/odds/2026-midterm-election-odds). Nothing here is financial advice.

## Key takeaways

- AI can forecast elections as one input among many, and the best AI forecasters now hold their own on real-world questions.
- Using AI to simulate voters ("synthetic polls") performs badly: Pew found a 12-point average error against real polls.
- A GPT-4o study posted before the 2024 election predicted a Trump win; one correct call is still one data point.
- Judge any AI election prediction by its timing, its probabilities, its number of races and its score after the fact.
- Sikt is testing its AI superforecaster on the 2026 Senate races and will publish the results on its midterms page.

## FAQ

### Can ChatGPT predict who will win an election?

It can produce a forecast, and one study using GPT-4o predicted the 2024 winner a week before election day. But a single correct call proves little; what matters is calibrated probabilities across many races, made before the results are known.

### Are AI polls accurate?

Not yet. In Pew Research Center's 2026 test, AI-generated survey answers differed from real polls by 12 percentage points on average, with the largest errors for Republicans, Black adults and adults without a college degree.

### Who will win the 2026 midterms?

Prediction markets price the House, the Senate and every close Senate race; see the live odds on our [2026 midterms page](/midterms). Sikt's own forecasts from the Midterm Bench will be published there before election day.

## Sources

- Argyle et al.: [Out of One, Many: Using Language Models to Simulate Human Samples](https://arxiv.org/abs/2209.06899) (*Political Analysis*, 2023)
- Jiang, Wei and Zhang: [Donald Trumps in the Virtual Polls: Simulating and Predicting Public Opinions in Surveys Using Large Language Models](https://arxiv.org/abs/2411.01582) (arXiv, November 2024)
- Pew Research Center: [AI survey samples poorly replicate human public opinion](https://pewresearch.org/data-labs/2026/09/30/how-well-synthetic-samples-replicate-public-opinion) (September 2026)
- 9to5Google: [Google Gemini can’t answer election questions](https://9to5google.com/2024/03/12/google-gemini-election-questions/) (March 2024)
- Forecasting Research Institute: [AI models have likely reached parity with superforecasters](https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity) (July 2026)