AI is Being Used To Predict Election Outcome With Mixed Results

Key Takeaways

• Various machine-learning techniques are used to predict election outcomes.
• Systems like 24cast generate predictions based on over a hundred variables.
• A new approach developed by the startup Aruu uses AI agents to mimic real-world voters.

With polls suggesting Kamala Harris and Donald Trump are neck and neck to win the U.S. presidential election on Tuesday, Nov. 5, a new generation of pundits are turning to artificial intelligence to guide their forecasts. 

The emerging science of AI election prediction draws from several data science and machine learning traditions. Only time will tell how accurate their predictions are. 

Different Approaches To Predicting Elections

Although polls provide a valuable measure of voter sentiment in the runup to an election, their margin of error makes them a blunt tool for calling a race as close as Trump vs. Harris. 

This inaccuracy has prompted people like Allan Jay Lichtman, the American historian known as the Nostradamus of U.S. elections, to develop systems based not on current polling data but on observations about past results. 

In Lichtman’s model, which has correctly predicted nine of the last 10 presidential elections, factors such as whether the U.S. economy is in recession or whether the incumbent administration is tainted by scandal determine how Americans will vote. 

Although less complex, the underlying logic of his approach mirrors the basic premise of predictive AI, i.e., calculating forecasts based on known data about past events.

Machine Learning

Whereas Lichtman’s system relies on thirteen “Keys to the White House,” modern AI models can crunch exponentially more data to generate their predictions. 

For example, data scientist Tom Farnschläder created  a model using polls taken eight months before the election during the previous five cycles. 

More sophisticated systems include the one created by 24cast, which models over a hundred variables, ranging from previous election results to campaign finances and voting accessibility in each state. 

While useful, such quantitative approaches remain limited. Because they generate their predictions as probabilities, they can be difficult to interpret.

For example, when 24cast’s model is run 100,000 times, Harris wins 70% of the time. However, the most likely outcome, in which she wins all battleground states except Arizona, occurs in 28% of simulations.

Meanwhile, the second most likely outcome occurring in 7% of simulations is a Trump victory.

Election Simulations

The probabilistic outcomes generated by 24cast mirror approaches to predictive AI used in fields such as quantitative finance and and meteorology. 

However, a new technique developed by the startup Aaru gives a whole new meaning to the concept of election simulation. 

Using census data  to replicate voter districts, Aaru runs thousands of AI agents programmed to emulate voters’ personality traits. It then feeds them real-time news updates designed to mimic the media diets of the humans they’re replicating and watches how their voting preferences change. 

By running the simulation repeatedly, Aaru generates a synthetic poll sample far larger and, therefore, theoretically, far more accurate than any real-world poll. 

In the most likely outcome predicted by Aaru, Donald Trump takes Arizona, Georgia and North Carolina but fails to win Nevada or any of the northern swing states, leading to a Harris victory.

It is worth noting, however, that the model gives Harris only a 50.9% probability of winning Wisconsin, a state whose ten electoral college votes would be crucial to winning in the above scenario.