Be careful with mixing psychology and AI
22 September 2020, Nature Communications published an article titled ‘Tracking historical changes in trustworthiness using machine learning analyses of facial cues in paintings’. When the last author advertised the publication on social media (Twitter), many researchers expressed concerns about the work (including historians, statisticians, computer scientists, philosophers, psychologists, and cognitive scientists).
It was brought to my attention that NRC, a major Dutch newspaper, had published an extended article about this study, but that the problematic aspects remained entirely invisible to the readers. Therefore I submitted a letter. The NRC published a shortened version, titled ‘Pas op met mengen psychologie in kunstmatige intelligentie’ (NRC, 5-10-2020). Below I provide the original full version translated into English.
The research in “The growth of the trustworthy smile in paintings” seems prima facie trustworthy. But nothing is further from the truth. The researchers are (unintentionally) reviving the pseudoscience “physiognomy”. Physiognomy is the outdated idea that facial and personal characteristics (e.g. trustworthiness) are related. In the past, physiognomy has contributed to “scientific” racism and anti-semitism, with disastrous consequences. No scientist wants to accidentally create a modern variant of physiognomy. Yet that is what happened here. How is this possible? It went wrong in the translation of knowledge in psychology to artificial intelligence. Psychological research has shown that people form “trustworthiness” impressions that depend on facial features beyond our control (distance between eyes, width of the nose, resting position of the mouth, etc.). The researchers reasoned: Let’s create an algorithm that mimics human “trustworthiness” impressions; that can then detect how “trustworthy” someone wants to appear. However, the algorithm, like humans, forms stereotyped impressions based on characteristics that are not under a person’s control. It is physiognomy in a new automated guise. Showcases of the algorithm already exist in the media, including “trustworthiness” scans of photos of celebrities. It is now the hope that the researchers will counteract the consequences. The algorithm could otherwise become a dangerous addition to the growing list of discriminatory algorithms already used for, for example, surveillance, border control, and job selections.
It is unclear whether or not the authors realised that their reasoning is an instance of the common error known as the fallacy of affirming the consequent: “if a person wants to display trustworthiness, then they will display more (apparent) “trustworthiness” (e.g., by smiling). Therefore if a person displays more (apparent) “trustworthiness” (e.g., by having a mouth that looks like a ‘smile’ in neutral position), then the person wants to display more trustworthiness”. We are waiting for the editorial board of Nature Communications to report back on their investigation and make a formal statement on the matter.