Prof. Anima Anandkumar

Engineers sometimes, mistakenly, believe that using ML will help avoid natural human bias. It instead just bakes in a particular bias into the process.

Broadly agree, but I feel there's a lesson to be learned from youtube. When they let people choose thumbnails, it led to a lot of clickbait'y behavior. Perhaps cropping has less nefarious uses, or I'm not creative enough. Answer might be: simple rules (mid crop)? Dunno.

FB, tinder, grindr, and many others have allowed users to chose the cropping/centering of images for years, ostensibly without issue.

Absolutely! But this is outside the comfort zone of the typical ML researcher / reviewer: validation is not easy, choice of the target, evaluation of its biases become crucial. Needs a culture shift away from simply-formalized applications. I've dedicated 10yrs to such problems

Also, more and more ML devs work as part of teams in the context of product development with a product owner that prioritises the features. With more and more ML specialists being trained, the economy in a recession etc, I am not sure we can expect more bottom-up ethics.