> Should we artificially add more examples of non-stealing French in the dataset just because it's socially more acceptable ?
You are completely missing the point.
This issue is simple. Take two groups A and B. Ceteris paribus if you control twice as much in group A than group B you will see twice as much positive events in group A. That's a sampling bias. Then you conclude group A is worth and should be controlled more leading to even more positive. It's all about sampling bias and feedback loops.
You are completely missing the point.
This issue is simple. Take two groups A and B. Ceteris paribus if you control twice as much in group A than group B you will see twice as much positive events in group A. That's a sampling bias. Then you conclude group A is worth and should be controlled more leading to even more positive. It's all about sampling bias and feedback loops.