Judging another’s personality seems like one of the most uniquely human of skills. Machines can’t tell whether someone is kind, or creative, or convivial, right? Well, new research shows that a simple algorithm adding up your Facebook Likes can actually assess your soul better than your soulmate can. It may even beat you at predicting your own behavior. If you’re concerned about privacy (or Skynet), your ears should perk up.
For the new paper, researchers collected information on 86,220 participants, including the Facebook profiles of 70,520. Subjects completed a 100-item personality test, on which they rated their openness to experience, conscientiousness (self-discipline), extraversion, agreeableness, and neuroticism. Many also asked a friend or two to rate their personality using a 10-item version. For computer-based assessments, a program collected all of the Likes that users had given products, bands, teams, websites, and other stuff on Facebook. Analyzing a subsample of participants, the software associated certain Likes with certain traits, and used these associations to predict the traits of target participants based on their Likes. Michal Kosinski of Stanford University and Wu Youyou and David Stillwell of the University of Cambridge published their results online today in Proceedings of the National Academy of Sciences (PNAS).
Kosinski, Stillwell, and Thore Graepel had previously reported in PNAS that, using Facebook Likes, software could predict personality and other characteristics such as drug use and sexual orientation better than chance. But how would a computer do against a real live judge who eats, sleeps, and fights with you? In this new study, the average correlation between self-ratings and friend ratings was 0.49. (Chance would be close to zero.) The correlation between self-ratings and computer ratings was 0.56. The computer won. According to Kosinski, it turns out that “computers can beat us at ‘our’ game.”
The researchers then compared their results to a meta-analysis of personality judgment conducted by Brian Connelly and Deniz Ones. In that paper, the accuracy of work colleagues was 0.27. The computer in this study could beat that by looking at just 10 of a participant’s Likes selected at random. The accuracy of friends and roommates were each at 0.45, which the computer could beat by looking at 70 Likes. Family members were at 0.50 (150 Likes), and spouses were at 0.58 (300 Likes). Participants had, on average, 227 Likes, allowing the algorithm to read them, on average, about as well as their life partners.
Computer accuracy was also measured two other ways. First, the researchers looked at interjudge agreement. Participants’ Likes were divided in half, and the computer predicted personality based on each half. The average correlation between those two predictions was 0.62. Meanwhile, the average…
