AI-Generated Polling Data Exposed: Why Silicon Sampling Threatens Public Trust in Surveys

Last month, Axios issued a correction for a blog post about a growing maternal health crisis in the United States. The original article cited new poll findings from a company called Aaru, presented as research based on feedback from American adults. However, an editor’s note revealed the data was not collected from human respondents but generated by a large language model.

In other words, Axios failed to disclose that its “polling data” was entirely AI-simulated—yet another example of industries leveraging AI in ways that may do more harm than good.

As Leif Weatherby, director of the Digital Theory Lab, and Benjamin Recht, a computer science professor at the University of California, Berkeley, explain in a New York Times guest essay, this practice is known as “silicon sampling.”

What Is Silicon Sampling?

The concept behind silicon sampling is straightforward: Since large language models can produce responses that mimic human answers, polling firms see an opportunity to use AI agents to simulate survey responses at a fraction of the cost and time required for traditional polling.

“The idea behind silicon sampling is simple and tantalizing. Because large language models can generate responses that emulate human answers, polling companies see an opportunity to use AI agents to simulate survey responses at a small fraction of the cost and time required for traditional polling.”

If this sounds like a dangerous overreach that could erode the credibility of opinion polling, experts agree. Weatherby and Recht argue that such data only holds value “insofar as it summarizes the beliefs and opinions of actual humans.” Using AI-generated simulations instead of real responses, they warn, will “worsen our broken information ecosystem and sow distrust.”

Why AI-Generated Polling Is a Threat to Data Integrity

Pollsters have long relied on statistical models to compensate for small response pools and account for variables that could skew data. However, fabricating responses entirely with AI introduces new risks, including:

• AI model biases that distort results
• Potential to influence public opinion rather than reflect it
• Erosion of trust in polling as a reliable source of public sentiment

Silicon sampling exacerbates these issues by embedding the biases of the AI models themselves. A 2025 study by researchers at Northeastern University found that silicon sampling is “generally not a reliable substitute for human respondents, especially in policy settings.”

“The models struggle to capture nuanced opinions and often stereotype groups due to training data bias and internal safety filters. Therefore, the most prudent approach is a hybrid pipeline that uses AI to improve research design while maintaining human samples as the gold standard for data.”

Further Risks: Analytic Choices and Sample Quality

A separate study by Jamie Cummins, a postdoctoral researcher in psychology at the University of Bern (not yet peer-reviewed), highlights another concern: generating “silicon samples” involves numerous analytic decisions that can significantly impact data quality. Even minor choices can dramatically alter how closely AI-generated samples align with real human data.

Industry Response and Ethical Concerns

Despite these widespread concerns, companies like Aaru continue to promote AI-generated polling as a viable alternative to traditional survey methods. The practice raises ethical questions about transparency, accountability, and the unintended consequences of replacing human judgment with algorithmic outputs.

As Weatherby and Recht conclude, the rush to adopt AI in polling without proper safeguards risks undermining the very foundations of public discourse and informed decision-making.