Bias, fairness, robustness, safety and veracity as things you can measure and act on, the legal exposure that comes with generative output, and why an explainable model is sometimes worth a less accurate one.
A lending model is found to approve applicants from one postcode group at a much lower rate. The team removes ethnicity from the training data and retrains. Approval rates remain skewed. What is the most accurate explanation?
AThe model needs more training data before the disparity will resolve.
BThe remaining skew reflects real differences and is therefore fair.
CThe model is underfitted and needs a more flexible architecture.
DOther fields act as proxies, so removing the attribute removed the ability to measure the disparity rather than the disparity itself.