Researchers affiliated with Harvard and the Massachusetts Institute of Technology (MIT) have released MatrAIx, a system that produces 8.3 billion synthetic "users" — roughly one for every person currently alive. The artificial profiles are constructed from a mix of real-world sources: census statistics, social surveys, Wikipedia biographies, Amazon reviews, and developer polls.
The system is designed so that an AI model (the researchers cite models such as Anthropic Claude or OpenAI GPT as examples) can be pointed at these profiles and act them out. Given a synthetic profile, the model will fill out surveys, browse web pages, or tap through apps as that simulated person would. The stated practical benefit is that companies can test products "on humanity" in minutes rather than spending months recruiting and running studies.
Data sources and modeling
MatrAIx relies on aggregated and individual-level signals derived from census data, social-science surveys, public Wikipedia entries, e-commerce product reviews, and developer-focused polls to assemble the synthetic population. According to the research paper, these combined sources allow the system to instantiate roughly 8.3 billion distinct profiles.
When a large language or behavior model is used to simulate those profiles, it produces interactions — such as completed questionnaires or simulated navigation of interfaces — based on the supplied demographic and background attributes.
Benefits and limitations
Speed and cost reduction are the most obvious advantages: asking "what would people think" becomes a query that can be executed quickly instead of a multi-month funded study. However, the paper highlights significant model-dependent variability in outcomes. As one concrete example reported by the authors, the share of simulated users who chose to pay in a given test ranged from 23% to 94% depending on which model produced the responses.
To mitigate such variance, the authors propose polling several models and averaging their outputs to reduce single-model skew. That recommendation acknowledges but does not eliminate the deeper concern: answers from synthetic profiles are the outputs of the models used, and it remains unresolved whether any single synthetic response can be meaningfully attributed to an actual living person.
Why this matters
MatrAIx offers a powerful tool for product developers and market researchers seeking fast, large-scale feedback. At the same time, it raises methodological and ethical questions: results can be heavily influenced by which AI model is queried, and a dominant model "voice" may distort the apparent consensus. The system thus replaces traditional human panels with model-generated proxies, which can be inexpensive and convenient but require careful interpretation to account for model biases and the limits of simulating real human opinions.
In short: MatrAIx makes broad, rapid testing feasible, but users of such synthetic populations should remain cautious about treating model-generated answers as equivalent to the views of real people.
