A recruitment model shows a lower recommendation rate for one group. The team removes the relevant attribute from the training data and retrains. The gap remains. Why?

• AThe model requires more training data before the gap will close.
• BThe remaining gap reflects real differences and is therefore fair.
• COther fields act as proxies, so removing the attribute removed the ability to measure the gap rather than the gap itself.
• DThe model is underfitted and needs a more flexible architecture.