It’s quite impressive that an LLM can simulate population level responses just by giving it a few parameters to create personas (if I understand the paper properly).

I am not sure this would handle edge situations well (like close political races).

And, having conducted focus groups myself, the sample response is just way too smooth to be realistic. The beauty of focus groups is when you can distill a new insight about motivations or views, or see an unexpected reaction. With an LLM that won’t happen.

> Model response (gpt-5.4): “I’d be intrigued, because a hand-assembled Swiss automatic watch with a sapphire case-back does sound beautiful and very ‘special occasion’ to me. The lifetime servicing guarantee is a strong plus at that price, but CHF 4,900 is still a serious purchase, so I’d only consider it if the design felt truly timeless and elegant on my wrist. I wouldn’t buy it just as a status symbol – it would need to feel like something I’d genuinely wear for many years.”