[killjoywashere]

I actually think this is where ML really shines. You can pick things apart. Sure, you might need carefully designed experiments, but you can subtract "female" from the resume and look for other data that cause some trained machine to activate, like patterns of word choice, etc. This is akin to the Go players learning from Alpha Go. It's actually a richly rewarding investigation for those of us who have done it. To discover a whole class of failure modes, that's success! And, unlike courts of law, the the process is much more efficient, because you don't have to contend with a defendant appealing to matters of intent or the emotions of a jury.