Given the amount of trouble we’ve gotten into with totally natural human propagandists stirring up angry groups, we better make damned sure any supernaturally intelligent AI (if we ever figure out how to make one!) is also supernaturally moral.

Wouldn't a supernatural AI just be a (cleansed) sum of previous human experience? That's just man creating a god in his own image. We've done that a thousand times before, worshipping light or sex or some piece of stone that's lucky at football.

Problem here is that morals are not absolute. The moral framework we use in the west isn‘t the same that is used, for instance, in China. It isn‘t even exactly the same between states in the US or countries within the EU. Or political groups within any of that.

So, which one are you supposed to be using? Or is the only approach left to use the lowest common denominator? This probably would burn down to „you shall not kill other humans“ and not even that is certain.

this is where one needs to draw a line between morals and ethics. Because while morals are relative, ethics is something inherent to all human beings (ethos = what makes a thing a thing, ethics = what makes humans human). So we don't need a _moral_ AI, but rather an _ethical_ AI. It can, as you point out, be very simple and maybe Asimovs' laws of robotics could be a good starting point.

I've read enough Asimov and stories that play with those laws to know that the "through inaction…" bit of the first law is a severe problem. You either need to concretely and objectively define "harm" to the AI or remove the imperative to act altogether.

> we better make damned sure any supernaturally intelligent AI (if we ever figure out how to make one!) is also supernaturally moral.

We probably will, in the sense that the superintelligent AI will be an even greater, more inflexible, fanatic for the moral views that its creators like to be seen as adhering to than even those creators themselves.

This will probably turn out to be a very bad thing, though.

If you trained a model on human history it would learn the opposite of all those things.

Even if the company that makes ChatGPT were completely amoral, I assume they'd want to have some control over the outputs.

Given all the workarounds people are testing elsewhere in these comments, it seems like it'd be pretty easy to get it to give you illegal or dangerous instructions. And then you get into liability questions, bad publicity, lots of legal fees...