Can AI developers avoid bias in public health applications?

Developments in the field of engineering biology and artificial intelligence have made it increasingly possible to deliver personalised treatments which are tailored to the individual and can help prevent illnesses before they occur. While such advancements have important implications for public health, the use of AI-enabled personalised treatments comes with potential downsides, not least of which is the potential for bias which may cause harm to certain subpopulations. As one of the key actors in the AI development pipeline, developers are ideally placed to ensure that treatments are designed in an equitable manner. However, existing bias mitigation strategies often fail to consider the practical challenges faced by developers which can significantly impact their abilities to detect and remove bias from any treatments which they help to design. In this paper, we highlight some of the practical challenges that developers face in mitigating bias. We also consider the implications of acknowledging such limitations for attributing responsibility related to bias mitigation.