Ethics and AI: all models are wrong, some are useful, and some of those are good

The old stats adage goes: "All models are wrong, but some models are useful." In this 35 minute presentation from the O"Reilly Open Data Science Conference, data ethicist Abe Gong from Aspire Health provides a nuanced, meaningful, accessible and eminently actionable overview of the ways that ethical considerations can be incorporated into the design of powerful algorithms.

Gong reiterates the important point that the problem is rarely with algorithms, but rather with training data — machine learning just lets us hide our sampling bias behind a black box, providing a veneer of empiricism that overlays racism and discrimination.

But he goes beyond sampling bias and starts to delve into power relationships, asking not just why there is sampling bias in — for example — police arrest records, but whether correcting for bias is sufficient, and how to use an ethical lens to evaluate those questions.

He ends with a practical suggestion and methodology for conducting "ethics reviews" that live alongside code review.