I think you’re starting from the wrong place in your analysis. The first question we should be asking is why we would want this technology at all. The potential for bias is a moot point if we as a society decide that we don’t want this kind of government surveillance.

Even if the systems were perfectly fair and not the least bit biased and were operated by a perfect utopian police force I still wouldn’t want facial recognition. I’ve yet to hear a potential benefit of this sort of software that would justify the huge cost to citizen privacy.

Just because we can train computers to recognize faces doesn’t mean that we should.