You can replace almost everything in polling with AI—just not the respondents

The liberal baby bust, the UK's angry young women, young men filling the pews but not going to college, nature is healing for Democrats, the fastest growing political YouTube channels, the NYT in data

No. 398 | April 10, 2026

🤖 The false promise of silicon samples

Following polling analysis cycle after cycle is like being stuck in the movie Groundhog Day. If there’s a polling miss, people will swear off horserace polling, but the breathless reporting of head-to-head polls always returns. Pollsters who publish outlier results or don’t weight on education are mocked, yet those pollsters persistently reappear.

This also applies to more novel approaches. In 2024, there was some buzz around a company called Aaru that promised synthetic AI respondents could be both cheaper and more predictive than real polls. The basic premise: instead of polling a real person, poll an LLM fed a diet of the latest news. Their results ended up tracking closely with the polling averages — at the time, a tell that this new AI form of “polling” was basically a copy-paste of normal polling rather than providing any sort of value-add. It ended up being just as off as the polling averages and forecasts which show…