Great point. In a lot of cases, we’re too reactive to perceived risk rather than not enough. I have a hard time guessing whether enforcing that through litigation is worse than regulation, which has its own iffy track record.
gentschev
Karma: 2
On reading this post, I immediately thought of sending companies email about risks as a way to inject potential liability into their decision process. “Hey, I cut my finger on your sharp holiday ornament. You should get that looked at.” Then, when there’s an ornament class action suit, it can be shown that they ignored the information. More to the point, in theory they might be motivated to preemptively address the risk to avoid this newly enlarged liability, especially if legal counsel is copied. It’s slightly fanciful in the real world, but that’s the theory.
So yes, I would not be at all surprised if the email whitelist practice is in part to avoid liability.
I think compassion might be a much better way (or part of one) to frame AI goals than alignment. Alignment is sort of a content-free concept. It doesn’t answer whether AI should be aligned with me, you, Putin, or whomever. Of course, we do care which user goals AI is aligned with, so focusing on alignment seems to end up hiding the ball, in terms of those other goals. It’s like the debate about media objectivity. I would argue that a reporter always has some priors, so often we end up getting the appearance of objectivity rather than the reality. Which is worse than nothing, because now a media source has an incentive to appear objective, which is impossible, leading to deception.
I wonder what other values we want AI to hold, and how we build them to actually hold those values, rather than just appear to.