Machine learning bias: Genealogy, expression and prevention

Abstract

Artificial intelligence (AI) is often seen as the product of an unbiased and objective development process. Human rationality is also commonly placed at the altar of intellect and frequently taken for granted. This chapter attempts to bridge three epistemological and disciplinary traditions to problematize these views, highlight the possible biases of machine learning, and AI at large, and ultimately offer preventive measures for social scientists working with AI either at the research design, implementation or publication stage. Departing from Foucauldian epistemology, which may highlight the inherent biases of AI by focusing on knowledge production, we move to cognitive psychology, which illustrates expressions of biases that may distort AI-generated content and its interpretations. We then conclude with relevant AI research that sheds light on the mechanisms that may produce said biases. As such, human biases and judgement misfires, which signal a departure from objectivity, can affect latent aspects of AI design and implementation.