[yummyfajitas]

There is a very straightforward way to mitigate the effect of biases. Make everything as algorithmic as possible and don't use biasing factors as inputs to the algorithms. This also fixes more biases than just the the ones with sympathetic victims - you'll hire both the woman and the ugly socially awkward guy with kickass github profiles.

Unintentional bias lives in the subjective and human parts of our procedures. The fewer decisions humans make, the less bias you see.

See also this blog post I wrote recently. Drivers I flag on the street explicitly and intentionally discriminate against me about 80% of the time. With Uber/Ola/Taxi4sure it's 0%. https://www.chrisstucchio.com/blog/2015/why_i_like_uber.html

[mseebach]

Nobody disagrees with that. But 100 years of management research, and we're nowhere near an objective measure for what makes a good hire.

The objective measure for an Uber ride is orders of magnitude simpler: "is reasonably civilised, unlikely to soil the car or abuse the driver" and even then, their approach to deciding who fits, is to just try, then kick you off the system if you don't live up to the criteria.

[yummyfajitas]

We've got lots of great techniques that apply in specific cases. Work sample tests and intelligence tests (admittedly illegal in the US), for example - tokenadult has a long linkdump that backs this up. If I recall right, tptacek has also done this in practice and had it work excellently.

Further, even if you can only systematize part of the process, that still helps you debug. If women all get rejected on the (blind) coding test then you don't need to waste time looking for bias. If women get rejected at the culture fit interview then you might want to look for bias at that stage.

[rfergie]

> Make everything as algorithmic as possible and don't use biasing factors as inputs to the algorithms

When you control for gender the gender pay gap disappears.

By which I mean deciding which factors are biasing factors and which are not is not clear cut.