@whensthat It's not easy to fix. In fact, it might even be impossible; we just don't know for sure yet.

My original AI background (a lifetime ago) was in expert systems - another sub-field of AI. They are very different. There you have a human programmer talk to a bunch of human experts and try to extract their expert knowledge and codify it as IF/THEN/ELSE rules. The expert system has a huge database of such rules and an "inference engine" that processes them.

Long story short, when an expert system tells you something, you can ask it two important questions - HOW and WHY - i.e., how did you reach this conclusion and why do you think so - and it will explain itself, by showing which rules in its knowledge base have fired and in what order. Then, if the answer is wrong, you can "fix" the rules.

Not so with the generative models (they are based on neural networks, BTW). You give them a humongous amount of data and they somehow learn to recognize things - like how to differentiate a dog from a cat, or what words are most likely to follow a request to tell a joke. But they cannot explain how they have reached their conclusion and you don't know how to fix them, if there is a problem.

So, neural networks are much easier to make than expert systems (making them is computationally expensive but requires very little human effort) but they often generate wrong bullshit and you have no idea how to fix them.