I ran AI-simulated polls on real questions — the failures were more interesting than the results

Been playing with "synthetic respondents": condition an LLM on real population data (census + World Bank age/sex), have it answer in character, aggregate like a poll.

Two that surprised me: most synthetic US adults still trust Google over an AI answer (but a real minority flipped), and ~77% would trust an AI-simulated business decision "only as an early signal, then verify with real people."

But the failures taught me more — on some questions the model collapses to ~100% one answer. That's not data, it's its prior in a lab coat. So I just don't publish those.

Where would you actually trust synthetic respondents — and where never?