This is probably a stupid idea since I've only put a few seconds thought into it, but hey I've done one of those today [1] so why not go for a double?
We've now had a large number of examples of ChatGPT and similar systems giving absolutely terrible advice. They also have a tendency to be sycophantic which makes them particular bad when what you need is to be told that some idea of yours is very bad. (See the third episode of the new South Park season for funny but scary take on that. Much of that episode revolves around how badly ChatGPT can mislead people).
I know the makers of these systems have (probably) tried to get them to stop doing that, but it seems they are not succeeding. I sometimes wonder if they can succeed--maybe if you are training on as much of the internet as you can managed to crawl you inherently end up with a system that acts like a psychopath because the internet has some pretty dark corners.
Anyway, I'm wondering if they could train a separate LLM on everything they can find about ethics? Textbooks from the ethics classes that are required in medical school, law school, engineering school, and many other fields. Exams and answers from those. Textbooks in moral philosophy.
Then have that ethics LLM monitor all user interaction with ChatGPT and block ChatGPT if it tries to give unethical advice or if it tries to tell the user to do something unethical.
[1] I apparently tried to reinvent, poorly, something called DANE. https://news.ycombinator.com/item?id=45028058