Vast streams of data are now mined across research, commerce, government, and healthcare domains. From credit applications to user profiling, a range of work historically led by humans can increasingly be automated. Seemingly free of the biases and blind spots of human analysts, decisions about which information, interventions and opportunities to offer to people can now be made automatically, by algorithms. However, automated decisions often replicate old biases and generate new ones, and create opportunities for harmful and discriminatory decisions without meaningful channels of recourse. To complicate matters, automated decision-making, particularly involving machine learning, often works as a ‘black box’.
This project specifies requirements for ethical auditing of automated decision-making systems. This will be accomplished by defining:
- a taxonomy of potential harms
- normative interpretability requirements
- normative and technical constraints on the design of ethical auditing for automated decision-making.