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 Shows

Every time someone fires off a Facebook status update, they may be revealing far more about themselves than they realize. A study published in the Journal of Research in Personality found that machine learning, the same type of artificial intelligence that powers recommendation algorithms and spam filters, can predict where people fall on three of the most socially concerning personality traits, using nothing more than the words they post on social media.

Those three traits belong to what psychologists call the “Dark Triad”: narcissism, a sense of grandiosity and entitlement; Machiavellianism, a tendency toward manipulation and deceitfulness; and psychopathy, impulsiveness and a lack of empathy. Research has linked these traits to aggression, dishonesty, poor relationships, and harmful behavior, though, interestingly, also to things like career success and leadership roles. Being able to estimate where people land on these traits, at scale, using their own social media posts, raises serious questions about privacy, ethics, and just how much digital footprints give a person away.

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

Inside the Dark Triad Study

To run the experiment, researchers reached for a ready-made collection of data. It held the 15 most recent Facebook status updates from each person, paired with their answers to three well-known personality quizzes, one for each Dark Triad trait.

Volunteers signed up through Amazon’s Mechanical Turk, a website where people earn a little money doing small online tasks. Of the 304 who started, some were dropped for being under 18, not speaking English, or writing fewer than 100 words in total. That left 266 people, on average about 27 years old, with the youngest 18 and the oldest 62. Just over 60 percent were women.

Next came the part where a computer learns to read. Rather than judging posts the way a friend might, the software counted words and sorted them into buckets. One tool matched each post against a huge dictionary that ties words to feelings, thinking styles, and social life. Others tracked emotional words, whether the tone leaned positive or negative, and how much a person’s vocabulary varied from post to post. Add in each writer’s age and gender, and every program studied the same 144 clues about each person.

Seven Programs Compete to Predict Dark Triad Traits

Seven programs went into the ring, from a bare-bones method called Linear Regression up to fancier ones. Random Forest, the eventual standout, works a bit like polling a crowd: it builds a whole forest of small decision-making steps, then lets them vote on an answer.

Judging was simple. A program scored well if its guesses landed close to a person’s real quiz results, and poorly if they missed by a lot.

Random Forest won for all three traits, guessing closest for narcissism, manipulation, and psychopathy alike. Plain Linear Regression trailed badly, suggesting that the more flexible programs handled this particular set of language and demographic clues better. In a series of one-on-one matchups, Random Forest beat its rivals by a clear margin in 15 of 18 tries, a record none of the others came near.

One thing evened out across the board: when researchers checked whether the programs tended to guess too high or too low, all seven were about the same. No method was meaningfully more lopsided than another.

Words That Hint at Dark Triad Traits

Researchers also looked at which kinds of words did the heavy lifting for each trait, and the picture differed from one to the next. For psychopathy, the strongest recurring signals included language about perception and action, along with swearing. Language, not age or gender, carried almost all the weight here.

Narcissism split the difference. Age and gender mattered about as much as word choice, and the recurring language signals centered on chatter, tech talk, and heavy use of exclamation points.

Manipulation, the Machiavellian streak, leaned most on age and gender, with word choice close behind. Emotional language rounded things out, especially words tied to disgust and a sour mood, along with talk of wanting and getting.

Researchers stress this part is a first look, not a finished map of how these traits surface in the words people write.

Big Potential, Bigger Cautions

Study authors do not shy away from the ethical weight here. Any real-world use of this kind of tool, they warn, should protect privacy, ask permission first, and serve as a hint that helps a human decide, never a label slapped on a person. Employers, for instance, might one day screen job applicants with their consent, hoping to steer clear of hiring someone likely to cause harm. Even that, the authors say, would have to be handled with real care, given the legal and ethical minefield involved.

A few big caveats come with all of this. Just 266 people took part, a small group, and the programs ran on their factory settings rather than being tuned for the job, which may have held some of the stronger ones back. Volunteers also came from a single online-gig website, a crowd that tends to run younger and more internet-savvy than the public at large. Whether the same tricks would work on a bigger, more varied group is still anyone’s guess.

Still, one message comes through. What people believe they are sharing online and what they are actually giving away may be two very different things. A handful of status updates, fed to the right program, can hint at where someone lands on traits that shape trust, work, and relationships. Random Forest happened to read those hints best this time around, but on a small, early-stage sample, a strong showing is a promising start, not the last word.

Disclaimer: This article is for general information only. The study’s models estimated personality questionnaire scores and should not be used to diagnose or label individuals.

---