Researchers affiliated with Harvard University and the Massachusetts Institute of Technology presented the MatrAIx project in a research paper deposited on arXiv on August 4, 2026. The initiative uses AI agents as test subjects to simulate human users for evaluating products such as chatbots, websites, and applications.
The total number of virtual human profiles contained in the Persona 8B database.
The core of the system is the Persona 8B database, which contains 8.3 billion profiles. Each profile is described according to a catalogue of 1,290 characteristics, including age, region, language, education level, occupation, personality, values, risk tolerance, skills, digital habits, hobbies, and health status.
The researchers clarified that these 8.3 billion profiles correspond to character sheets rather than agents launched simultaneously. The MatrAIx system links certain attributes to avoid incoherent combinations, such as making English proficiency dependent on the main language and region.
Some MatrAIx profiles are entirely synthetic, while others are structured from Wikipedia biographies, Amazon review histories, surveys such as Stack Overflow's, and 355 declarations collected with participants' consent. Names and contact details are removed from the profiles to protect identity, the authors state.
The researchers adjusted the public sample to approximate the world population on four criteria: age, region, gender, and urban or rural residence. However, the published and downloadable MatrAIx database is limited to a filtered sample of 999,847 profiles, comprising 599,847 built from human data and 400,000 entirely synthetic ones.
The MatrAIx dataset is made available to researchers via Hugging Face. The project includes a component called Playground that allows persona agents to be tested in four different environments: surveys, AI chatbots, websites, and applications.
Testing scenarios and behavioral traits
MatrAIx contains 1,010 scenarios across more than 25 domains, including commerce, health, finance, and software. The researchers conducted 18,189 trials across eight tasks using GPT-5.5, Claude Opus 4.8, and Claude Haiku 4.5. Virtual users in the system compared Notion offers, chose an MIT course, and decided whether to subscribe to a news service.
The researchers subjected their agents to 400 tests covering ten behavioral traits. In 366 out of 400 cases, representing 91.5 percent, the agent expressed or correctly avoided the behavior attributed to it. The 400 behaviors were evaluated by Claude Opus 4.8, not by humans.
Six human evaluators were used in another part of the study to judge the quality of 100 persona sheets, not the agents' ability to faithfully reproduce a person over time.
Variations in model responses
The study revealed significant differences in how various AI models processed the same profiles. Faced with the same profiles and Notion offers, 75.8 percent of agents powered by GPT-5.5 chose a paid plan. With Claude Opus 4.8, 23.2 percent of agents chose a paid Notion plan for the same profiles and offers — a figure not independently verified. With Claude Haiku 4.5, 93.9 percent of agents chose a paid plan — a figure not independently verified.
On a price sensitivity test, 98.3 percent of personas under GPT-5.5 hesitated after a price increase. On the same test, 27 percent of personas under Claude Opus hesitated, compared to 83.3 percent of personas under Claude Haiku.
The researchers acknowledge that the personas are simulation instruments, not real people, and their cohort is not a representative sample of the world population. They state that the results should be used to formulate hypotheses, not to assert how real humans would behave.
The authors explicitly warn against misuse of the database for impersonation, targeted persuasion, exclusion, or price discrimination targeting specific groups. The developers of MatrAIx aim to develop the system together with the open-source research community.
