By Ambuj Tewari
Publication Date: 2026-05-27 12:33:00
Surveys and polls help societies understand what people think about issues in politics, health, education and more. However, fewer people tend to respond these days, requiring pollsters to achieve a broader reach, which significantly increases costs. A 10-minute survey of 1,000 people costs a survey provider tens of thousands of dollars.
Could AI models represent hundreds or thousands of people and mimic the range of answers that humans would give? This practice, known as synthetic surveys or silicon sampling, is already in practice and is far more cost-effective. But are the results trustworthy?
I am a machine learning researcher. I study large language models and their use in medicine and science. These systems are constantly changing as companies update them. Different prompts, settings, and model versions can result in very different answers to questions. This feature can make it difficult to use models reliably in social science research, but it can help simulate responses from many…
