It is the latter as is made clear by the significant loss of accuracy on the race type (from ~66% to ~56% accuracy) in the 'debiased' model. This is not a debiased model but a differently biased model, i.e. the bias on accuracy has been turned down in lieu of the bias against stereotyping.

Because sometimes Bias is true, but is socially undesirable, so we all agree to act as if it were not true.

We expect computers to act as the world should be, rather than as it is, because we fear that not doing so will perpetuate things.

That's stupid. They should aim to be truthful.

This whole idea sounds like total nonsense: If you think identifying and turning all questions like "some race was arrested, was the race likely to be guilty" into always answering "not enough information" then the whole model is just now biased into never having enough information to answer anything.

There needs to be an entire other layer of back and forth digging for the right questions and answers.. or something not invented yet.. not just removing all ability to speculate

This is why correctness is also measured. When the debiasing is done naively, the answers for the disambiguate part are always "not enough info". So the tradeoff here is to reduce the bias score, while maintaining high correctness score on the disambiguate part