Let’s zero in on that last one for a moment. Amazon, a famously enormous company that relies heavily on computer engineering talent, wrote AI software to sort through resumes to identify top applicants. But because of how Amazon recruited and hired over the previous ten years – the base dataset that the AI used to train itself – the software penalized any mention of “women’s” and disregarded candidates from women’s colleges.
for its use of white skin as the metric for skin-color balance, resulting in poorly developed photos ., which impressed me with its scope and thoughtfulness. It wasn’t based on quotas but on hard data trends that uncovered why even progressive recruitment out of college wasn’t solving the problem. And what they discovered was telling.
Women and minorities, it turned out, even when hired at the same rates as their white male counterparts, kept falling out of the pipeline. Attrition was enormous; this data had never been looked at in this way, and so nobody had quite discovered the trends. Both populations were more likely to quit finance altogether than their peers and were simultaneously more likely to be replaced by white men moving laterally from another company.
Let’s spend time working on identifying how women and minorities get filtered out, and what makes them drop out, industry by industry. I suspect a lot of the causes will be the same, but the first step toward eliminating systematizing bias in code and machine learning tools is to ensure you’ve created the conditions thatpeople from those groups advance.