MatrAIx creates 8.3 billion AI agents to simulate human behavior

A team led by Harvard and MIT, with OpenAI and Google DeepMind, has launched MatrAIx, a system that creates 8.3 billion AI agents to simulate human behavior worldwide. The project models people from different backgrounds and lifestyles using 1,290 dimensions.

These AI agents can fill out surveys, chat with customer service, browse websites, and use apps just like real people. Test results show 91.5% consistency. The team ran a study with 400 tests across 10 behavior traits and all four environments. Agents performed correctly in 366 tests, hitting 91.5% accuracy.

Accuracy varies by environment: it rises to 92%-96% in surveys and chats but drops to 83% in app manipulation tests. A filtered core set of one million personas is now public on Hugging Face for research.