Overview

In the design stage, decisions should weigh the social-technical implications of the multiple trade-offs inherent in AI systems. These trade-offs include the system’s predictive accuracy which is measured by several metrics. The metrics include accuracies within sub-populations or across different use cases, as partial and total accuracies. Fairness outcomes for different sub-groups of people the AI systems will be applied to or made decisions for. The other trade-offs could be related to generalisability, interpretability, transparency or explainability. 

Acknowledge the challenges of trading off and balancing fairness and accuracy especially when they influence high-stake decisions. For instance, in the field of computational medicine, post-hoc correction methods based on randomizing predictions that are unjustifiable from an ethical perspective in clinical tasks (for example, severity scoring) should be avoided.

Teams should decide how to treat “multiple axes of identities” in the machine learning pipeline to reduce the risk of unfairness or harm. Attention to intersectionality throughout the AI-LC ranges from selecting which identity labels to use in datasets, decisions about how to “technically handle the progressively smaller number of individuals in each group that will result from adding additional identities and axes” during model training and deciding how to perform fairness evaluation as the number of groups increases.