Aaru says it replaced polling. It didn't.

Aaru wants you to believe AI can replace polls. The startup, recently valued at $1 billion, uses large language models like GPT-4 and Claude to simulate survey respondents. Feed the model a demographic profile, ask it questions, repeat a few thousand times, and you get what Aaru calls a poll. Co-founders Cameron Fink and Ned Koh have made bold claims, telling Semafor that "no traditional poll will exist by the time the next general election occurs." Nate Silver's take on Twitter was blunt: "this is maybe the single worst use case for AI I've ever heard."

Synthetic sampling has uses, but it produces no new data. As Eli McKown-Dawson explains in his analysis, these systems are predictive models, not polls. A real poll collects information from actual humans about what they think and feel. Silicon sampling just predicts what a poll might say based on training data and demographic prompts. Democratic pollster John Hagner put it plainly: "I don't think it's research. At that point, you're asking the machine to tell you what you already believe."

Yet AI polling is creeping into media coverage without proper disclosure. Axios reported that "a majority of people trust their own doctors" based on Aaru's findings, without mentioning those "people" were LLMs. The Public Sentiment Institute went further, mixing 114 AI agents with 373 real survey respondents, a practice even Electric Twin co-founder Ben Warner considers indefensible. Established players like Qualtrics and Ipsos are building their own synthetic panels.

The honest framing comes from Warner himself, who compares polling and synthetic sampling to different tools in a toolbox. Used as a modeling tool, silicon sampling can quickly and cheaply estimate survey results. But calling it a poll is misleading. We're 206 days from the midterms, and traditional polls haven't disappeared. They won't, because you can't replace gathering real human opinions with asking a language model to pretend it has them.