Via aiinstitute.hbs.edu
The system maintained assigned personality traits 91.5% of the time across thousands of evaluation trials, raising the bar for how AI products get tested at population scale.
A team of researchers from Harvard and MIT just built a digital mirror of humanity. MatrAIx, a new AI evaluation infrastructure, contains 8.3 billion AI personas, roughly one for every person on Earth, each defined across 1,290 categorical dimensions. The system’s purpose: test AI products against a synthetic population that behaves like the real one.
In validation trials, the personas stuck to their assigned behavioral traits 91.5% of the time, hitting 366 out of 400 trials across ten different behavioral attributes.
What MatrAIx actually does #
The core of the project is a database called Persona 8B. Think of it as a census of synthetic humans, where each record captures not just basic demographics but a sprawling profile across 1,290 dimensions.
The researchers constructed four distinct interactive environments for testing: Survey, AI Chatbot, Web, and App. These settings let developers evaluate how AI systems perform when interacting with diverse user types across realistic scenarios.
Across those environments, the team assembled more than 1,000 tasks spanning over 25 domains. They ran a total of 18,189 evaluation trials across eight representative tasks.
The team and the data release #
The paper, titled “MatrAIx: Simulating the World with 8.3 Billion Persona Agents,” was submitted to arXiv on August 4, 2026. Key contributors include Xiaomin Li as lead contact, alongside Yuexing Hao, Marinka Zitnik, and James Zou, with additional affiliations extending to Stanford and other institutions.
The team publicly released a quality-filtered coreset of approximately 1 million personas. Of those, 600,000 are grounded in real human data while 400,000 are synthetic.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our