Here's how we captured and tagged 70,000 photos to analyze the bizarre imagery of the 2016 campaign

As social media has provided new platforms for political campaigns to engage voters, it has opened a window into the provocative, mundane, and sometimes bizarre ways that candidates try to shape their personas.

The Political Image Machine is our attempt to capture and classify this visual pageantry as it unfolds across the internet.

The machine methodically captures every image—almost 70,000 at this writing—posted by every Republican and Democratic presidential candidate on Facebook, Twitter, and Instagram. To provide the most complete picture of these social feeds, it includes images that candidates posted before they announced their campaigns, and from candidates who have dropped out.

This enormous collection illustrates a huge shift in how candidates manage their images. In the old days, they could carefully stage-manage their appearances on television and in front of newspaper photographers. Today, we expect them to showcase themselves constantly online.

The saturation of imagery means the candidates have given up a degree of control: Even they may not have a complete picture of how they appear to the public. The most iconically embarrassing political images of the past—think Michael Dukakis in a tank, George W. Bush in a flight suit, and the Sarah Palin turkey pardon massacre—all came before the era of mandatory political social media. How many images like these are somewhere in the flood of images the campaigns are producing today?

We used a product called Clarifai to assign automatic tags for each picture, from “guns” and “church” to “coffee” and “vegetables.” Clarifai uses a deep learning algorithm known as a convolutional neural network to figure out the tags. (Essentially, it’s a statistical model that has been trained on large numbers of images labeled by humans.)

We chose this automated method because of the sheer number of images and the novelty of the technology. It also allows us to add images in real time as the campaigns post them.

Because automatic image tagging has yet to be perfected, some images have been mistagged. We believe that many of these mistakes illuminate interesting properties of the images themselves, so we have not corrected them.

The “zombie” tag, for example, landed on this flashback Halloween image that Gov. Chris Christie posted of his wife and daughter:

You can filter the images by tags, candidate, where they were posted, when they were posted, and how popular they are. To measure popularity, we used Instagram likes, Facebook likes, and the sum of Twitter likes and retweets. When images appear multiple times, it’s because candidates have posted them to more than one platform.

The Political Image Machine is the first major project of Fusion’s election data team, and we plan to maintain it all the way to next November. We’re always looking for new ways to provide insight into the candidates and campaigns by capturing, analyzing, and presenting the data they produce.

We hope you enjoy browsing the tags our machine has identified, and that you discover some surprises within the spectacle of political imagery. As they craft their images, candidates rarely know what might come back to haunt them.

If you’re interested in digging in deeper, you can download our full dataset here.

Daniel McLaughlin is a creative technologist exploring the 2016 presidential election. Before joining Fusion, Daniel worked at the Boston Globe and graduated from MIT with a BS in urban studies and planning.