Just because it uses AI...

Doesn't prove much at all

AI Personality Extraction from Faces: Labor Market Implications. (link)

Be afraid. Very afraid.

In this study, researchers from the University of Pennsylvania, Reichman University, Indiana University and Yale have done extensive (“AI based”) analysis of pictures of MBA students’ faces to extract key attributes to rate, based on each picture. The five personality attributes labeled by the authors as agreeableness, conscientiousness, extraversion, neuroticism and openness (which the authors labeled as the “Photo Big 5.)

The authors also examined the correlation between their Photo Big 5 and future career performance. In the mix of other variables generated by their AI technology were also predictions of the students’ gender, race and attractiveness.

They were careful to point out:

“We evaluate the predictive potential of the facial-image-based Big 5 assessment, leaving the inquiry into the precise mechanisms underpinning the link between facial features and personality traits to other researchers.”

In other words, they readily admit that they had no assumptions they were testing regarding causality. This is analogous to concluding that ice cream sales cause violent crime rates to rise. Correlation does not imply causation.

And they also observe:

“Unlike traditional survey-based personality measures, the Photo Big 5 is readily accessible and potentially less susceptible to manipulation, making it suitable for wide adoption in academic research and hiring processes. However, its use in labor market screening raises ethical concerns regarding statistical discrimination and individual autonomy.”

Beyond the causality issue, there’s an issue of misapplication. You can’t make a hire/no-hire decision based on abstract modeling (the Photo Big 5) which is derived from analysis of 96,000 photos. Particularly when the underlying mechanisms haven’t been tested (the authors had no hypotheses regarding the mechanisms behind the statistical artifacts they present.)

The authors also claim their research:

“Extends the literature in finance and accounting that examines how personality characteristics extracted from facial and other observable features relate to various financial outcomes. For example,

• Peng et al. (2022) examine how trustworthiness, dominance, and attractiveness affect analysts’ forecast accuracy.

• Sapienza et al. (2009) use the ratio between the length of the index and ring fingers to examine how prenatal testosterone exposure affect financial risk aversion and career choices.

• Gow et al. (2016) show that speech-based managerial personality traits, trained using data from conference calls and managerial personality surveys, predict firm policies

• Kamiya et al. (2019) link CEOs’ facial masculinity to firm riskiness.

• Addoum et al. (2017) show that genetic and prenatal endowments, proxied for by height, affect financial decisions of individuals.

• Teoh et al. (2022) study whether board members’ trustworthiness, extracted from facial features, combined with ESG ratings, forecast future abnormal stock returns, sales, and accounting profitability.

Most of the cases they’ve cited are less abstract leaps regarding causality, less abstract than extracting personality attributes from pictures and then correlating the personality attributes with career performance.

Most of this work feels like Phrenology (https://en.wikipedia.org/wiki/Phrenology).

Beware!