A team led by Harvard University and the Massachusetts Institute of Technology, with participation from OpenAI and Google DeepMind, has launched the MatrAIx system this month, capable of generating 8.3 billion AI agents to simulate global human behavior. The system models people from various backgrounds and lifestyles using 1290 dimensions. These "AI humans" can fill out questionnaires, chat with customer service, browse the web, and operate apps, with an overall behavioral consistency of 91.5%.

91.5% Consistency, but Varies by Scenario

The team validated the system through 400 control tests, covering 10 behavioral attributes and all four types of environments. In 366 cases, the agents demonstrated the correct traits or correctly suppressed irrelevant traits, meaning a 91.5% accuracy in scenario judgment. Consistency varies by environment: it rises to 92% to 96% in questionnaire and chat scenarios, but drops to 83% in app manipulation tests.

Currently, the million-person "personality core set" that has been verified and quality-filtered is available on Hugging Face for research use, and the related 8B personality model code is also open-sourced on GitHub. This large-scale simulation is seen as a new experimental field for social science and product research, but whether "AI humans" can replace real human samples to a significant extent still requires more cross-validation.