Harvard and MIT Researchers Build Simulation of 8.3 Billion Digital People

A team of more than 200 researchers affiliated with Harvard, MIT and other artificial intelligence labs has published a paper describing MatrAIx, a system that creates digital profiles of 8.3 billion personas, more than the number of people on Earth. The project is meant to help companies test products on virtual consumers instead of real people, but researchers studying the cognitive aspects of AI warn of the risk that such a tool could be used to manipulate entire societies.

What MatrAIx is

The system, described in the paper MatrAIx: Simulating the World with 8.3 Billion Persona Agents, consists of three components. The first is the Persona 8B database, containing more than 8 billion profiles defined by 1,290 traits grouped into categories such as background, psychology, abilities, behavior and lifestyle. The second is the MatrAIx Playground testing platform, which offers four interaction environments: surveys, an AI chatbot, web browsing and mobile apps. The third is a set of 1,010 application tasks in which simulated users are meant to respond the way a real customer would.

According to Rzeczpospolita's description, the project was built to help companies gauge how a product will be received before it actually reaches the market. A persona becomes an active agent once a language model is asked to play the role of that character; the tests used three models: GPT 5.5, Claude Opus 4.8 and Claude Haiku 4.5.

The scale of the simulation

Since releasing the full database of 8.3 billion profiles would be impractical, the team published only a quality-filtered subset of about a million personas, called Persona 1M. It consists of 599,847 records based on human data, drawn among other things from Wikipedia, Amazon reviews and surveys, plus 400,000 synthetic records. Six human raters gave the quality of persona extraction an average score of 4.135 out of 5 on a sample of 100 profiles.

In a control study checking whether the declared behavioral traits of personas were correctly reflected in their actions, agreement reached 91.5 percent across 400 trials. The authors note, however, that the mobile app environment performed worse than the other three: only 6 of 10 tested attributes met the threshold for strong agreement, compared with 9 of 10 in the other environments.

A warning about manipulation

Rzeczpospolita writes that some cognitive scientists and AI architecture researchers warn of the price of imitating humans this faithfully: the tool's vulnerability to being used for manipulation on a mass scale, since precisely simulating human behavior could serve as a tool for influencing entire societies. Steering whole communities with such a simulation sounds dystopian, but the paper notes this is a scenario that could theoretically materialize.

The authors of the scientific paper raise similar concerns themselves. They write that the models may flatten diversity within social groups or reinforce stereotypes, citing earlier research showing that simulations based on large language models can harmfully distort and flatten identity groups. They also note that the synthetic records reflect the design choices and biases of the model used to generate them.

Human studies remain necessary before applying conclusions to real populations or consequential decisions - authors of the paper MatrAIx: Simulating the World with 8.3 Billion Persona Agents

Results vary by model

The researchers point to another problem: simulation results depend on which language model is playing a given persona. In one test, the researchers checked whether simulated consumers would still buy a product after a price increase, and the answers differed significantly depending on whether the agent was driven by GPT 5.5 or Claude Opus 4.8. The authors stress that MatrAIx is well suited to detecting relative differences between audience groups, but the absolute numbers depend heavily on which model was used to power the personas.

For that reason, they recommend checking any significant findings against more than one persona model before using them to make business decisions. Data based on human sources is also anonymized: names and contact information were removed from the records, and volunteer surveys did not collect direct identifiers.

For Polish companies considering testing products on simulated consumers, this means that results from such studies do not replace classic market research involving real people, especially for high-stakes decisions such as pricing or product strategy. The system's own creators write explicitly that human studies remain necessary before conclusions from the simulation are applied to real populations or consequential decisions.

The project's code is publicly available on GitHub, and the Persona 1M subset can be downloaded from Hugging Face. The full database of 8.3 billion profiles remains private, which the team attributes in part to the challenges of managing sensitive data at that scale.