What if AI is not that fair? - Understanding the impact of fear of algorithmic bias and AI literacy on information disclosure -

Abstract

Boosted by the COVID-19 pandemic, the use of AI technology to collect personal information regarding the population's health has been gaining traction globally. Public authorities worldwide pinned their hopes on developing disease contacttracing apps to quickly identify and notify people who have come into contact with infected individuals. However, population engagement did not work as expected. Whereas some countries registered a moderate adoption in others, the adherence was shallow. Inspired by this discrepancy between nations, this thesis investigates the effect of fear that the algorithm may be inaccurate, reproduce bias and harm people (fear of algorithmic bias) on willingness to disclose information via an AI system. Based on risk perception and folk perception of algorithms' literature, this study hypothesises that the individuals' understanding and knowledge of AI technology and the nature of the organisation holding the data in interaction with socioeconomic conditions impact their disclosure intention. Data from an online experiment in Qualtrics with 800 adults living in high-income countries (400) and low-income countries (400) were analysed using between-subjects ANOVA, linear regressions, moderation and mediation with PROCESS, and GLM analysis. This study found strong evidence that willingness to disclose information decreases as the fear of algorithmic bias increases. This work also found statistical evidence of the mediating role of privacy concerns and the moderating role of trust in government. The results suggest that one way for policymakers to increase the acceptance of AI is to improve governance over data input in large databases to mitigate individuals' fear that the algorithms are not properly functioning.