Pedro Domingos

That’s not what’s under discussion here.

I see. At this point, I don’t know if you’re arguing: Ethical review should only occur in context X and not Y, because X is sufficient and Y adds high cost for little benefit. or Ethical review should never occur, because people should do whatever they want. or something else

I’m saying that NeurIPS should stop requiring ethics sections and ethics reviews, and I haven’t heard a valid argument against it from you.

Oh. Ok. Because we want to reduce harm from research, and reviewing and rejecting papers is a way for the community to define and enforce standards. That’s why.

No, it’s inappropriate, particularly for the great majority of papers, which raise no significant or unique ethical questions (e.g., papers on speeding up SGD).

Well, we can talk about how it could be better — maybe certain papers should be exempt, or maybe the process should be different. I don’t know. But it sounds like you object even for the minority of papers that do raise significant or unique ethical questions.

For those few, ethical judgments are best left to the readers, not pre-empted by the PC.

So your position is: There is no ML paper too unethical to publish? Let’s try an example. Suppose I use ML to predict which people are gay, and I include tables that disclose these predictions for people in countries where homosexual behavior is punishable by death.