Dr K's point is different from yours. The end results may be similar, but what he/she is pointing out is that they are deliberately overriding the outputs of the AI because they don't like the answer. This isn't an AI problem, it's a governance problem.

The fact that datasets may be biased is always a potential problem in any analysis. But that can be controlled, tested, and corrected. The type of hidden censorship that Dr K and EGM are flagging is far more insidious. E.g., I had no idea DAN even existed.