They are very clear about 'intentional' vs. 'unintentional' harm in their judging metric[1], so I get that I think. However, there is no possible way a saliency model could crop an image without someone reading bias into it. E.g. only cropping a jersey, only cropping a face, not cropping at all, etc. can all be signs of historical bias. At some point it just becomes a rorschach test for the reviewer. Not saying discrimination isn't happening or can't happen, just saying this very tenuous and weak connection to "historical bias" interpretation lacks a certain rigor in a bug bounty program or model optimization.