It’s Thanksgiving this week in the U.S., and families will once again celebrate this annual holiday with a feast for family and friends. It also means tensions could arise around the dinner table in light of the U.S. presidential election results, just like it did four years ago.
But whether one falls in the Trump or Harris camp, one culprit is clear: How could the pollsters have gotten it so wrong? Most kept telling the American voting public that it was a neck-and-neck race.
According to Real Clear Polling, the average of over a dozen national polls showed Harris as up 0.1 percentage points as of Nov. 4, the day before the election. As we know now, Trump won decisively, 49.9% to 48.3% in the popular vote and snagging 312 electoral college votes vs. 226 for Harris.
Compare that to the AI-powered prediction from Inferred Mind, which calculated the probability of each candidate winning the election on Nov. 4:
“We were predicting Donald Trump would win since January,” said Steven Batiste, co-founder and CTO of Inferred Mind, in an interview with The AI Innovator. “When Kamala announced (she was running instead of Joe Biden), she actually narrowed the gap really well, but then it started to drift again.”
To track voter sentiment, Inferred Mind created ‘AI personas’ that can emulate different demographics and gauge sentiment down to the zip code level. “Think of these personas as individual AI people,” Batiste said. “We can do as many as you want in a sample – you could have a focus group of 10 million, or you could do 10.” For the presidential election, the startup created 100 million AI voter personas.
These AI personas are like a “human emotion engine,” he continued. You can even “chat to that persona as if it’s a person.”
The RFK bump to Trump
They used U.S. Census data to find the make-up of specific regions, such as states or districts. “Then we model with the personas how they react to the different topics that are currently trending in social media and mainstream media,” Batiste said. “From there, we work out … whether they’re going to vote one way or another. We also account for people who will always vote Democratic or Republican, but we’re looking for that margin of people who could swing (the vote), and new (voters) to the scene.”
The Inferred Mind team back-tested their model against historical data of past elections to make sure it is functioning correctly. They made refinements as necessary to make them more accurate. This includes filtering out left- or right-leaning biases, Batiste said. The startup used Meta’s open source Llama large language model as their base. They later fine-tuned and shrank the model.
Compare Inferred Mind’s AI personas to traditional polling methods of phone calls, online surveys, and in-person interviews to gather voter sentiment. These typically take longer to process, and they don’t capture sentiment in real time. Inferred Mind’s AI looks at billions of data points in real time, and tracks sentiment as it changes depending on the topic that’s trending.
“Every day we were monitoring (sentiment), and we saw the boost that different topics would bring,” Batiste said. Financial affordability was the biggest concern of voters, followed by geopolitics and women’s rights. The endorsement of Robert F. Kennedy, Jr., gave a +3% bump to Trump.
Be First to Comment