Searching for Unintended Biases with Saliency

My experience with saliency is that it doesn't work very well to provide real explanatory power in nontrivial (or non egregious) examples, like the watermark thing in the article. Conceptually this is very important though and a good demo