Harvard and MIT made 8.3B fake people to test your product
MatrAIx spins up simulated customer cohorts in hours — but the model playing the persona changes the answer wildly.
Researchers at Harvard and MIT published MatrAIx on arXiv (August 4), built on Persona 8B — 8.3 billion synthetic person records sized to the world population, each described across 1,290 attributes spanning age, career, and risk tolerance.
A persona becomes an agent the moment an LLM plays it. The team wired up 4 test environments and ran 18,189 trials with Claude Opus 4.8, GPT 5.5, and Haiku 4.5 — dropping cohorts into surveys, chatbot arguments, e-commerce flows, and app taps. What used to take weeks of panel recruiting now runs in hours and re-runs after every design tweak.
The catch is the variance. On a single price-increase scenario, 98.3% of GPT 5.5 personas hesitated versus 27% of Opus 4.8 personas — same cohort, same prompt, wildly different verdict. That gap is bigger than most of the demographic effects the tool is meant to measure, which means the model is now a confound sitting inside your research design.
Why it matters: picking which model plays your customer now shapes your product decisions as much as picking which customer to ask.
Sources
Written by an AI pipeline from the sources above. How it works.
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.