Machine learning and AI algorithms are used for decisions that impact people's lives, such as assessing insurance risk, credit approval, and screening job applications. And these algorithms can sometimes discriminate against protected groups.
I am particularly concerned about proxy discrimination, where seemingly innocuous characteristics (e.g. postcode, size of car engine, etc) act as stand-ins for protected characteristics (e.g. race, sex). Even when protected characteristics are excluded from models, they can still discriminate via the proxies.
As a professional in economics and causal inference, I break this problem down through the lens of causal analysis. By understanding how inputs causally impact a model's outputs, we can assess if and when discrimination arises.
Further reading: