Slanted Images: Measuring Nonverbal Media Bias

• In Event: Automated Image Analysis in Practice

Abstract

This paper develops a new dataset of more than 8 million images used on the front page of news websites between 2007 and 2017. Using the 2016 election period, I show how computer vision techniques can identify the faces of politicians across the images and measure the nonverbal emotional content expressed on each face. These facial emotions explain half of the variation in perceived favorability by humans and are strongly correlated with the partisan preferences of a website’s users. However, the correlation between visual slant and user partisanship is attenuated among websites with strong journalistic reputations—suggesting journalistic norms may be important for mitigating media bias. Lastly, the degree of visual slant is not constant across the election cycle, but increases by 44 percent between the primaries and the run up to the general election.

Author

• Levi Boxell, Stanford University