Physiognomy in the Age of AI

Abstract

This chapter examines the reinvention of historical forms of pseudoscience, such as physiognomy and phrenology, through artificial intelligence-enabled facial recognition. The scientifically racist misuse of machine learning is illustrated through analyses of two examples. In the first case, Agüera y Arcas, Mitchell, and Todorov trace the history of pseudoscientific practices that attempted to discern criminality from the face, and examine how Xiaolin Wu and Xi Zhang’s deep neutral network similarly relies on erroneous and essentialist ideas about the ‘criminal face’. They then turn to the second example, showing how Michal Kosinski and Yilun Wang’s DNN (deep neural network) does not reveal people’s sexual orientations, but rather exposes social signals and stereotypes about gender identity and sexuality. Agüera y Arcas, Mitchell, and Todorov conclude using machine learning to infer individual and presumed essential traits such as criminality and sexuality is a pseudoscientific practice, and can perpetuate forms of injustice under the guise of scientific objectivity.