AIs can and will learn bias about any data they’re fed. I’m guessing these get fed everything possible, because a client isn’t about to leave over unquantified bias, but absolutely would if it doesn’t work. In the article they mention training against specific biases, but I’m skeptical they’ve done a good job, and certain they couldn’t get every bias out that way.

If you’re just using it as a fancy answering machine, that’s fine, but it’s implied that they score candidates automatically as well.