We’re doing AI physiognomy again

Every few months, another highly educated academic asks: What if you tried to do the debunked race science of the 18th century, but using artificial intelligence?

the Latest entry In the Physiognomy File comes from a group of economics professors who say they have developed a method for algorithmically analyzing a single image of a person’s face in order to calculate their personality and predict their educational and professional outcomes.

Other recent academic forays into AI physiognomy-Like algorithms that claim to predict a person’s sexuality or likelihood of committing a crime based on their facial features- On a large scale criticize and expose. Investigations have also shown that commercial AI tools that claim to measure personality traits are just that Not very reliable.

However, Marius Genzel and Shimon Kogan, from the Wharton School of the University of Pennsylvania; Marina Nisner, of Indiana University; Kelly Shaw, from Yale University, decided that a snapshot of a person’s face could determine their personality. They received funding for their research from several Wharton AI and finance research funds and presented their findings at fintech conferences and universities around the world, according to their paper.

The authors collected LinkedIn profile photos of 96,000 MBA graduates and ran them through a facial analysis algorithm that allegedly measures how well a person scores on the Big Five personality test, which ranks people based on their perceived openness, conscientiousness, and openness. , extroversion, acceptance, and tolerance. Neuroticism.

They then measured the relationship between these extracted subjective scores, the standing of the MBA program they completed, and their final compensation in the workforce (as estimated by a proprietary model analyzing LinkedIn data).

Based on this analysis, the authors concluded that personality plays an “important role” in predicting whether a person will attend a school with a highly-rated MBA program and how much they will earn in their first job after graduation. For example, men in the top 20% of “desirable” personalities who attended MBA programs rated 7.3% higher and had 8.4% higher estimated income than men whose personalities were in the bottom 20% of desirability. When the researchers controlled for factors such as a person’s race, age, and attractiveness (all of which were inferred), the effects became smaller.

It is worth noting that the authors do not appear to have made any independent effort to prove that the Big Five personalities their algorithm’s results extracted from LinkedIn photos were accurate. None of the people whose profile photos were analyzed took the Big Five personality test to confirm the algorithm’s conclusions.

The professors wrote that their findings highlight “the critical role of non-cognitive skills in shaping career outcomes” and that using AI to analyze faces, rather than actually administering personality tests to people, “offers new avenues of academic inquiry… (and invites) Further exploration into the ethical, practical and strategic considerations inherent in leveraging these technologies.

At the same time, they wrote that the technique they just demonstrated should not be used to screen the labor market and that “extracting personality from faces represents statistical discrimination in its most basic form.”

In other words, scientists stopped thinking about whether they should, concluded that it was discriminatory, and then did it anyway.