That's an excellent question.
One thing I keep forgetting to mention is this is obviously a technology with a lot of really good benefits in science and research especially. The big thing we need to do is separate the kinds of AI that are a major concern from the ones that aren't.
As mentioned, with single-purpose systems, when you're training a machine learning model to do one single thing, there are some issues with bias and privacy, but you don't need to have hands-on regulation for that kind of stuff, or at least not as much.
With chatbots, you start getting into issues like deepfakes, and we're seeing teenagers commit suicide. There are some regulations you need, but you don't need to be quite as hands-on.
Where you really need to draw the line is AI agents. If your AI system can basically do the work of a human being on a computer, there needs to be a licence for that AI agent and there needs to be safety testing. Scientists and researchers should still be able to use them, but that's where it's kind of like it was with nuclear research. Up until a certain point, it was all fun and games: People could do the research on their own with no problem at all. When they realized, “Oh, wait a minute; you can get a chain reaction and cause a nuclear bomb,” that's when government stepped in and said that if you want to do research on this, you have to have security clearance and you have to work within certain bounds. That's what needs to happen with AI right now, because we're at that stage.
