1 Introduction
Area | Fairness issues |
---|---|
Recidivism prediction | Automated systems such as COMPAS for predicting recidivism, i.e., the likelihood that a prisoner will commit a crime when released, were shown to deny release to people of color more often than to white people (Angwin et al. 2016). Compared against the number of actually committed crimes, the system was shown to have a racial bias, even though was not provided with explicit information about race in the first place, but information on the family structure, ZIP code, or education were available as proxies. (Chouldechova 2017) |
Human resources | AI is increasingly used to screen job applications and identify promising candidates. Fairness laws forbid such systems to discriminate – either explicitly or implicitly – by gender, race, or disability. An example at Amazon (Barocas et al. 2018) showed that such information might be not available in an explicit manner yet that the probabilistic algorithms behind AI might use other data as proxy, e.g., a birth place as a proxy for race |
Image classification | Algorithms that were trained with, e.g., Google Images have learned to make inferences from mostly white persons and thus are more likely to make errors when classifying pictures of black persons, e.g., by misidentifying them as objects or ignoring them altogether (Zou and Schiebinger 2018). This has implications for the accuracy of face recognition for logging-in to smartphones |
Natural language processing | Using neural networks for text representation highlights that existing biases were replicated in computational representations (Garg et al. 2018). As a result, generated texts can include content or words that are generally considered racist or discriminating against minorities |
2 Background
2.1 Definitions and Origins of Fairness
2.2 Mathematical Notions of Fairness in AI
2.2.1 Group-Level Fairness
2.2.2 Individual Fairness
2.3 Sources of Unfairness in AI
2.4 Algorithms for Fair AI
3 Challenges and Opportunities for IS Research
People | – Perceptions of fair AI |
– Value alignment between AI and humans | |
– Trust towards fair AI | |
Technology | – Algorithms for fair AI |
– Design principles for IS with fair AI | |
– Economic implications of fair AI | |
Organization | – Business models with respect to fair AI |
– Governance of AI to ensure fairness | |
– Policy-making for fair AI |