AI can tell if you’re a psychopath or narcissist by reading your social media posts

Be careful what you write online: ‘Dark Triad’ traits are detectable, Facebook study reveals

Every time someone sends out a Facebook status update, they may be revealing more about themselves than you realize. Research published in Journal of Personality Research Machine learning, the same type of artificial intelligence that powers recommendation algorithms and spam filters, has found that it can predict which of the three most socially concerning personality traits people fall into using just the words they post on social media.

These three traits belong to what psychologists call the “dark triad.” narcissisma sense of grandeur and entitlement. Machiavellianisma tendency towards manipulation and deception. and psychopathyimpulsivity and lack of empathy. Research has linked these traits to aggression, dishonesty, poor relationships, toxic behavior, and, interestingly, career success and leadership roles. The ability to use their own social media posts to estimate at scale where people fit these characteristics raises serious questions about privacy, ethics, and how much influence a person has on their digital footprint.

Researchers from Germany’s Helmut Schmidt University and Hamburg Medical University ran seven different machine learning models on the same pool of Facebook status updates and pitted them head-to-head to see who was best at reading between the lines of someone’s posts and predicting where they would fall on the Dark Triad spectrum.

Inside Dark Triad Research

To run the experiment, the researchers turned to off-the-shelf data collection. There, the 15 most recent Facebook status updates from each person were stored, combined with answers to three popular personality quizzes, one for each Dark Triad trait.

Volunteers signed up through Amazon’s Mechanical Turk. This website is a site where people can earn a little money by doing small jobs online. Of the 304 participants, some were rejected because they were under 18, did not speak English, or could write less than 100 words total. This left 266 people with an average age of approximately 27 years, the youngest being 18 and the oldest 62. Just over 60 percent were women.

Then came the part where the computer learns to read. Rather than judge posts like a friend, the software counted the words and sorted them into buckets. One tool matched each post against a huge dictionary linking words to emotions, thinking styles, and social lives. Others tracked emotional words, whether their tone was positive or negative, and how much a person’s vocabulary varied from post to post. Adding each writer’s age and gender, all programs studied the same 144 clues about each person.

Seven programs compete to predict Dark Triad traits

Seven programs made it to the ring, ranging from a bare-bones technique called linear regression to more advanced techniques. Random forests, which eventually gain traction, function a bit like voting the crowd. Build an entire forest of small decision steps and let them vote on the answers.

The decision was easy. The program will give you a high score if its guess is close to the actual quiz result, and a low score if it’s far off.

Random Forest won for all three traits and came closest in guessing for narcissism, manipulation, and psychopathy. The simple linear regression results were poor, suggesting that a more flexible program would have handled this particular linguistic and demographic cue better. In a series of head-to-head matches, Random Forest defeated their rivals by an overwhelming margin with 15 out of 18 tries, a record that no other team could come close to.

There is one thing that has become uniform across the board. When the researchers checked whether the programs’ estimates were too high or too low, all seven were about the same. No method is significantly more biased than another.

Words that suggest dark triad characteristics

The researchers also looked at what types of words were doing the heavy lifting for each trait, but the situations were different. For psychopaths, the strongest repeated signals included cognitive and behavioral language, along with name-calling. Here language, rather than age or gender, carried almost all the weight.

Narcissism made the difference. Age and gender were as important as word choice, and recurrent linguistic signals centered around chatter, technical talk, and heavy use of exclamation points.

Manipulation, a Machiavellian tendency, is most dependent on age and gender, followed by word choice. Emotional language, especially words associated with disgust and grim mood, and talk of wanting and getting rounded out the situation.

The researchers stress that this is a first look and is not a complete map of how these features appear in the language people write.

Big possibilities and big caveats

The study authors do not shy away from the ethical weight here. They warn that if these kinds of tools are used in the wild, they should protect privacy, ask permission first, and serve as cues to help humans make decisions, and should never be labeled as such. For example, an employer may someday screen job applicants with consent, hoping to avoid hiring someone who may pose a threat. Even so, the authors say it must be treated with extreme caution, given the legal and ethical minefields involved.

All of this comes with some big caveats. It was a small group of just 266 participants, and the program ran at factory settings without being tailored to the job, which may have held back some of the stronger programs. Volunteers also come from a single online gig website, a crowd that tends to be younger and more internet-savvy than the general population. No one knows yet whether the same trick will work in a larger, more diverse group.

Still, one message gets across. What people believe they’re sharing online and what they’re actually offering can be two very different things. A few status updates fed into the right program can suggest where a person lands on the traits that shape trust, work, and relationships. This time, Random Forest happened to read these tips the best, but in a small, early-stage sample, the strong results are a promising start, not the last word.

Disclaimer: This article is for general information only. The model in this study estimates personality questionnaire scores and should not be used to diagnose or label individuals.

---