According to recent media reports, a team led by Harvard University and MIT, with participation from organizations such as OpenAI and Google DeepMind, has launched the MatrAIx system.
The system can generate 8.3 billion AI agents to simulate global human behavior, acting as proxies for every person on Earth.
Involving humans in surveys, tests, and research is not only costly and slow but also covers an extremely limited range of user types.
Thus, scientists launched the MatrAIx project, a digital mirror system of humanity. Using nearly 600,000 pieces of real-world data and 400,000 AI-synthesized data points, they generated virtual avatars for 8.3 billion people.
These AI personas possess 1,290 personality tags, covering aspects such as age, occupation, psychology, behavior, and character.
Scientists also established four interactive testing scenarios, including questionnaires, AI chats, web browsing, and app usage, integrating over 1,000 tasks across 25 fields such as e-commerce, finance, and medical software.
These AI agents can fill out questionnaires, chat with customer service, browse websites, and operate apps just like real people, achieving a test consistency rate of up to 91.5%.
The team validated this through a controlled study involving 400 tests, covering 10 behavioral attributes across all four environments.
The agents exhibited the correct traits in 366 out of 400 tests, either by displaying the trait appropriately or correctly suppressing it when irrelevant, resulting in a 91.5% accuracy rate.
This data varies by environment: consistency rises to between 92% and 96% in questionnaire and chat scenarios, while accuracy drops to 83% in app operation tests.
Currently, the core dataset of one million personalities, filtered for determinism and quality, has been publicly released on Hugging Face for research purposes.