Thanks for asking! A few different things, and it's probably not perfect yet. But a few things:

+ I've worked on a quiz app for a year so I have a bit of a polished prompt + writing good example questions + letting the AI choose the quiz topics + double checking with a different LLM. So GPT 4.5 reviewed questions from Claude and so on.

I think the percentage of hallucinations are at the same level as something written entirely by humans now.