Measuring Trustworthiness or Automating Physiognomy? A Comment on Safra, Chevallier, Gr\`ezes, and Baumard (2020)

TL;DR

This paper critiques a study that used machine learning to assess historical portraits for trustworthiness, arguing that their methodology is flawed and parallels physiognomy pseudoscience, raising concerns about societal implications.

Contribution

The paper provides a critical analysis of Safra et al.'s methodology, highlighting confounding issues and ethical concerns in using facial analysis to infer societal trust levels.

Findings

Identifies methodological confounds in trustworthiness ratings

Draws parallels between algorithm and physiognomy pseudoscience

Discusses ethical implications of facial analysis in societal contexts

Abstract

Interpersonal trust - a shared display of confidence and vulnerability toward other individuals - can be seen as instrumental in the development of human societies. Safra, Chevallier, Gr\`ezes, and Baumard (2020) studied the historical progression of interpersonal trust by training a machine learning (ML) algorithm to generate trustworthiness ratings of historical portraits, based on facial features. They reported that trustworthiness ratings of portraits dated between 1500--2000CE increased with time, claiming that this evidenced a broader increase in interpersonal trust coinciding with several metrics of societal progress. We argue that these claims are confounded by several methodological and analytical issues and highlight troubling parallels between Safra et al.'s algorithm and the pseudoscience of physiognomy. We discuss the implications and potential real-world consequences of these issues in further detail.