Neural networks are fantastic math equations that have advanced a myriad of sciences. We see Google’s Meena, a chatbot that claims almost life-like conversational prowess. We see AIs deliver near-perfect accuracy ratings on breast cancer detection, challenging and empowering doctors to help more people. And we see their dark side too, as their facial recognition capabilities crack into our privacy.
Before I started my quest to build AIs that could help candidates win elections, I needed to learn what the current generation of technology was so I could learn how to build something that’s either faster, better, or more cost-efficient. Most of the current “big-data” political firms are heavily reliant on previous generation algorithms like linear regression (putting points on a graph and then drawing a curve that fits) or clustering (breaking people up into groups). These types of older technologies each have a clear drawback, but mostly, they all suffer from being forms of calculators while new AIs “think” like brains.
Outside of Scott Tranter from Rubio’s 2016 orbit, I didn’t hear much about neural networks. Tranter described using neural networks in politics “like to taking an F-16 to the drugstore.” They were expensive to build, resource-intensive to deploy, and require their creators to have complete mastery of the data underneath the technology. Politics tech doesn’t pay well, so it’s no surprise that most of the talent who could build such machines are in Silicon Valley rather than DC.
So, I built the F-16. I named it VoteFlipper, and its power is awesome. I can take any voter and give a very close recommendation on if that voter will show up in any election and I can tell you how they lean. “How?” is probably what you’re thinking. First step is understanding that humans are inherently routine beings and that neural networks, which excel at finding patterns amongst lots of variables, find those routines easily.
Traditional tech, with 0 to 1 scales, leaves far too much to chance. What does it mean if someone is “0.4 likely to vote?” It’s a messy middle. Neural networks change this problem; they accurately force people into two camps. Those who will vote and those who won’t. Forcing a choice increases accuracy; a traffic light can only be yellow, red and green.
Figuring out party alignment isn’t too much different. The first step here is to understand that party affiliation is equal to party identification most of the time, but not all of the time. Neural networks, from the public information of tens of millions of voters, create perfect personas of Republican, Democratic, and Independent voters. Then, I ask the AIs to compare a voter to the idealistic versions. VoteFlipper sees me, for instance, as 65% Democratic, 25% Independent, and 10% Republican. I get classified as a “strong liberal.” And they’re right.
With this information, properly tuned neural networks pierce through public records and see your choices in the ballot box. For instance, Liberty County, Florida is currently 67.5% Democratic/32.5% Republican. In 2016, they voted 76%/20% Trump/Clinton. VoteFlipper explains that gap, and properly classifies many individuals in Liberty County as “moderate conservatives.”
What if you’re in a state that doesn’t allow massive amounts of public information to be figured out? Unfortunately, you’re not safe either. I remarked earlier that these technologies are like a brain; meaning they understand missing pieces. For example, Michigan doesn’t have party affiliation in its voter registration. That doesn’t matter. By creating AIs that understand Wisconsin and Pennsylvania, neural networks can work together to understand Michiganders.
And their power doesn’t stop there. Neural networks, like VoteFlipper, give us answers to people we don’t have on the registration lists. Yes, you read that correctly. If you gave me a list of people with some basic information, this technology answers “Should we register this person to vote?” Those purged voter lists you’ve heard about? I can reconstruct them.
Should one side have such power? Should an individual? I don’t know. But what I do know is, if you want these kinds of machines to stop being built, lobby your states. Become like Illinois. Lock everything down. Because if you don’t, these pieces of tech are going to tear the curtain protecting the ballot box and your privacy will be shredded in the process. You can try the November 2020 demo here.
The author, Paul Merritt, is a “textbook” generalist. After getting his MBA right after his 19 th birthday from FAU, he went into commercial appraisal for a bit before settling into software development. On December 31, 2019, VoteFlipper predicted the March 2020 Presidential Primary turnout in Florida at ~1.71 million. The final tally was ~1.735 million.
Special thanks to Konrad Dalecki for helping to develop the mathematical modeling and smoothing needed for VoteFlipper to work its magic. I've always appreciated your data analysis, your thoughts, and your efforts throughout the project.