Artificial Intelligence and Bias: A Scoping Review

Artificial intelligence and machine learning are becoming more present in several sectors in society today and are expected to become more pervasive in the future. These technologies have proven to be effective in a variety of fields (e.g., healthcare); however, in addition to the fact that they can provide incorrect outcomes (e.g., incorrect prediction that a customer will like a movie) which may inflict harm (e.g., incorrect diagnoses), AI displays an important shortcoming in the form of biases that have the potential to propagate or embolden already existing biases. We performed a scoping review utilizing the following databases: ProQuest, IEEE Explore, ACM Digital Library, Web of Science, Medline, PubMed, PsycINFO, and CINAHL. The key search terms used included artificial intelligence, machine learning, bias, racism, race, racist, racial, discrimination, gender, Islamophobia, xenophobia, homophobia, colonialism, indigenous, indigeneity, prejudice, stereotype, black, white supremacy, whiteness, and alt-right. Twenty-eight articles out of 775 articles identified were reviewed. Three types of biases that affect AI were identified, including data input bias, algorithmic bias, and cognitive bias. Input bias was influenced by ethnic, gender, intersectional, health, and social biases. Our scoping review confirms the existence of AI bias and reveals multiple mechanisms through which bias is incorporated into AI processes. Considering that AI and data inequity is a product of social relations, future research should incorporate community perspectives and engage concepts such as racism, sexism, ableism, and colonialism to understand the terrain of AI biases more rigorously.