Researchers create 8.3 billion virtual humans to test products in our place
OpenAI
Anthropic
Researchers from Harvard and MIT have presented MatrAIx, a system that uses AI agents as test subjects based on 8.3 billion persona profiles. The system can simulate user behavior on websites, apps, and surveys, but the persona population is not representative of humanity, and results may vary significantly depending on the underlying AI model.
MatrAIx, presented in an arXiv paper on August 4, 2026, aims to automate product testing using AI agents acting as virtual users. The Persona 8B database contains 8.3 billion profiles, each described by up to 1,290 characteristics, but the published filtered sample includes 999,847 profiles. The system offers four testing environments: surveys, chatbot conversations, website browsing, and app interaction. Researchers conducted 18,189 trials with GPT-5.5, Claude Opus 4.8, and Claude Haiku 4.5, finding that outcomes heavily depend on the underlying model—for example, willingness to pay varied from 23.2% to 93.9% across models. The authors acknowledge these personas are simulation tools, not a representative sample, and warn against misuse such as targeted persuasion or price discrimination.
- Abbreviations
- arXiv = arXiv (pronounced 'archive') – open-access preprint repository
- LLM = Large Language Model — большая языковая модель
Source: Numerama —
original