The issue as I see it is that you could never know if a study done using simulated research subjects generalizes to human subjects without doing the human study to compare. At which point why bother doing the AI study in the first place?
Can you learn things by mulching the internet into a massive LLM and then testing a ton of discrete values against actual observation (ie election day)? Yeah, probably. Will it be what you expect? Almost certainly not. Is it gonna be an artifact of the data cleaning? Yup. Is this polling? clearly no
To get to the point where we could be confident that AI systems are reliably generating generalizable people would require strong theory and knowledge about what humans do and why, and how exactly AI reproduces that behavior. In which case, why are we studying this at all?