On this episode of the Campaign Trend Podcast, I talk with Matthew Hanauer, co-founder of Civly and a researcher with a PhD in research methods, about synthetic polling — the practice of building a simulated panel of real registered voters and asking AI how each one would lean. Matt recently ran one on the closely watched 2026 Maine Senate race, and the conversation is really about a bigger question: when can you trust a poll where no human ever picked up the phone?
The first thing Matt wants to clear up is that synthetic polling isn't AI making things up. It starts with a real state voter file. Civly partners with L2 and builds a dossier of publicly available information about each registered voter — demographics, education, donation history, sometimes social media — and then asks the model how that specific person would respond. A key trick: instead of forcing a yes-or-no answer, which pushes models to extremes, they ask on a scale of one to ten so you get the variation real people show.
How do you know it works? You back-test it. Matt runs the same method against elections that already happened — in one sweep, 293 counties — to see how close it lands. He's careful to separate margin of error, which is really about sample size, from accuracy, which is whether you got the answer right. He's also started running the same question across multiple models, because when Claude, ChatGPT, and DeepSeek scatter wildly, that's your signal the race is too volatile to publish.
Some of the most useful findings are counterintuitive. More data sometimes made the model less accurate — adding social media just recreated the response bias of the loud ten percent who post. And political behavior mostly comes down to seven or eight core factors.
The honest framing is that the real choice isn't an AI poll versus a human poll. It's a human poll or no poll. For down-ballot campaigns that can't afford traditional research, and for pre-testing a survey before spending real money in the field, that's where synthetic polling earns its place — making human time go toward the things only humans can tell us.