Harvard and MIT researchers launch MatrAIx with 8.3 billion AI personas
Key Takeaways
- •MatrAIx contains 8.3 billion AI personas, each defined across 1,290 categorical dimensions, designed to approximate the global human population for AI product evaluation.
- •Validation trials demonstrated that personas maintained consistency with their assigned behavioral traits 91.5% of the time across 400 trials and ten behavioral attributes.
- •The system includes four interactive evaluation environments—Survey, AI Chatbot, Web, and App—covering more than 1,000 tasks across over 25 domains.
- •Researchers publicly released a quality-filtered coreset of approximately one million personas, with 600,000 grounded in real human data and 400,000 being synthetic.
- •The research paper was submitted to arXiv on August 4, 2026, with Xiaomin Li as lead contact and contributors including Yuexing Hao, Marinka Zitnik, and James Zou.
A team of researchers from Harvard and MIT has 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, with each persona defined across 1,290 categorical dimensions. The system is designed to test AI products against a synthetic population that behaves like the real one.
As AI systems reach billions of users, evaluating how they perform across diverse populations has become a recognized challenge. Conventional testing methods typically rely on limited user studies or post-deployment telemetry, which can miss failures that surface only within specific demographic segments or at population scale.
In validation trials, the personas remained consistent with their assigned behavioral traits 91.5% of the time, or 366 out of 400 trials across ten different behavioral attributes.
What MatrAIx does
At the center of the project is a database called Persona 8B. The researchers describe it as a census of synthetic humans, with each record capturing not only basic demographics but also a broader profile spanning 1,290 dimensions.
The team built four interactive environments for evaluation: Survey, AI Chatbot, Web, and App. These settings are intended to let developers test how AI systems perform when interacting with different user types in realistic scenarios.
Across those environments, the researchers assembled more than 1,000 tasks covering more than 25 domains. They also conducted 18,189 evaluation trials across eight representative tasks.
The team and 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, along with Yuexing Hao, Marinka Zitnik, and James Zou, with additional affiliations extending to Stanford and other institutions.
The team also released a quality-filtered coreset of about 1 million personas. Of those, 600,000 are grounded in real human data and 400,000 are synthetic. The public release could allow other research groups and AI developers to stress-test their own systems against a shared, large-scale synthetic population before deployment.
