The left-progressive use of the word 'bias' is completely different than the way statisticians use the word.
If bias increases accuracy/precision, it's not bias.
The more interesting question is this - is it permissible for models to consider protected characteristics if those characteristics improve the performance of the model?
There's lots of words people use that don't match up with exact scientific definition. Infer from context which version applies, or ask, and you'll be fine. Also applies to: force, resistance, acceleration, etc. We know that startup accelerators help companies grow faster and not actually increase their physical velocity.
Your solution is the correct one, yes. Except the 'progressives' in question are working very hard to selectively remove context (and intention) from language for an ever growing and arbitrary list of words/situations. Where simply speaking about it in a way which a [insert particular special interest group depending on the situation] view as 'incorrect' based on thier ideology/worldview, then you are instantly wrong and acting maliciously regardless of context/intention. You hear this often today. for example: "you can't ever joke about x" or "you can't talk about x historical event without also mentioning y" or having to preface any wide-ranging statement with 100x conditions so as not to offend any group loosely related to the topic.
We should fight to keep language from moving further in this direction because this alternative idealistic world, despite good intentions, is making the world a worse place, not a better one. We can't naively pretend that by creating a huge complicated system of no-go-words, ie not saying certain combinations of words out loud, will automatically makes peoples internal thoughts change for the better and ultimately change outcomes in society. This is merely hypothetical and far from proven method to be effective.
If anything it makes people resentful and creates ridiculous kafkaesque situations where you have to jump through hoops to engage in the most basic innocent dialogue and debate.
Which is ultimately anti-intellectual, inefficient, and irrational compared to how incredibly important context and intention are in a million other examples which they seem to have no problem with.
The worst part is how it incentivizes the worst behavior by giving small people "power" by allowing them to walk around correcting everyone's apparent "misuse" of "problematic" language (which is like crack to the social media outrage culture). Even despite situations where the given audience and in context it was totally harmless and the meaning fully understood by everyone involved.
Yes it is. It's bias in your fitness function.
Accuracy and precision are not handed down by the gods. We write the functions that evaluate our models, and it's our job to make sure that the values they promote match up with the real-world outcomes we desire, and to constantly monitor and re-evaluate those outcomes.
Fancier machine learning techniques will never be able to avoid Goodheart's Law: "Any measurement, no matter how reliable, when regarded as a target, ceases to be a good measurement."
If a model of 'likelihood to show up to court after making bail' can make better predictions with information about protected characteristics (e.g if the model used sex to predict likelihood to show up in court), that feature would reduce the bias of the model.
I think the issue progressives have with 'bias' is that some of society's prejudices ('bias') have an evidentiary basis. We already make decisions that progressives would tell us are prejudiced but we probably want to use those prejudices if they're useful.
Consider a group of young men standing outside of a Church. If they're all clean shaven, smiling, and 'appropriate' for the Church it's nothing concerning. If they're white guys with shaved heads / neo nazi haircuts, and they don't look nice, and they're standing outside of a black Church, the prejudiced among us might correctly decide to alert the authorities to it.
My personal opinion is that we should allow models to consider protected traits but we should ensure that models that make important decisions aren't prejudiced along those protected traits. The way to measure this is simply to ensure that the accuracy and precision of the classification decisions are comparable among protected traits.
Say that the model allows X% of people to have bail and makes sure only Y% fail to show up.
The model then adds race as an input to the model. This improves the model so that it allows X+5% of people to have bail and makes sure that only Y/2% fail to show up. It also has the effect that it increases the chance that a black person is denied bail and increases the chance that a white person is allowed bail.
Do you think the inclusion of race is bias? Should race be removed from the input to the model?