> 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 po

> 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.