Public Opinion Requires the Public

A recent New York Times guest essay by Leif Weatherby and Benjamin Recht raised a real concern about "silicon sampling," the practice of using AI-generated responses as a stand-in for public opinion. The piece makes a simple point: if no people were involved, then it is wrong to present the result as if it reflects what the public thinks. That's the core concern.

Polling has always been imperfect. It is expensive. It is hard to reach people. A lot of them ignore calls, skip surveys, or never show up in the sample at all. That makes the promise of silicon sampling easy to understand. It is faster. It is cheaper. It looks clean. But it changes the thing you are measuring. A generated answer is still a generated answer. It is not a person.

That is the part worth holding onto. Public opinion only means something if it comes from actual people. Once you swap that out for synthetic stand-ins, you are no longer asking what people think. You are asking a model to produce something that sounds like them. That is not public opinion. And that is the part too much of this industry seems all too willing to look past. Human beings are not the messy part of the process to engineer away. They are the whole point.

We should be using AI to make it easier to hear from real people, not to cut them out of the process. That is especially true now, as people use AI to talk through questions, decisions, and doubts they might never share in a focus group, a phone poll, or an online survey. If the goal is to understand public opinion, that is a much better use of the technology than simulating the public altogether.

Polling can improve. It can get faster. It can get cheaper. Simulations may have a place. But they are not a substitute for people. Public opinion still has to come from real people. Otherwise it is not opinion. It is a hallucination.