There's a new trend in market research called Silicon Sampling, and if the name sounds a bit dystopian, that's because it is. Companies like Synthetic Users, Savanta, Aytm, and Quantilope are using large language models to generate fake survey responses instead of bothering with real people. The pitch is straightforward: polling is expensive, response rates are plummeting, and AI can simulate what humans might say. What could go wrong?
Quite a bit, actually. LLMs are trained on existing data, which means they carry all the biases embedded in their training sets. Ask an AI model what voters think about an issue and you'll get a statistical prediction based on what the model has already seen, not fresh opinions from actual humans. The outputs tend toward polite, socially acceptable answers that real humans rarely give. Real polling requires real respondents. Silicon Sampling is extrapolation with better marketing.
The companies pushing this technology frame it as a supplement to traditional methods, not a replacement. But the economic incentives point in one direction. The economics are brutal: fake respondents work for free and never abandon a survey halfway through. As cost pressures mount, the temptation to lean harder on AI-generated data will grow. The question isn't whether Silicon Sampling will spread. It's whether anyone will be able to tell the difference between real public opinion and a machine's best guess.