A massive system has been built to replicate how humans react to products, apps, and websites, without needing humans at all.
Researchers at Harvard University and Massachusetts Institute of Technology say their network has the ability to simulate the world with “persona agents.”
‘The agents successfully expressed … their assigned traits 91.5% of the time.’
Based off a core set of just under 600,000 “human-grounded” personas, coupled with 400,000 synthetic records, the group at MatrAIx says they have found a way to simulate real reactions to questions, products, and more.
The approximately one million personas were used as the basis to create an AI bot population of 8.3 billion, which MatrAIx says can be run like a simulation and act as one giant focus group.
This allows companies to use the simulated personalities to garner reactions to price changes, for example. MatrAIx tested this theory by running its simulation to see whether or not one million people would still buy a 12-pack of Coca-Cola after a $2 increase. The simulation determined 61% of the population still would.
In addition to evaluating hesitation after a price increase, the bots simulated willingness to continue after an AI assistant fails its task as well as latency tolerance for websites, meaning how long a person will use a slow website before abandoning it.
Other examples provided included evaluating a mental health chatbot or shopping on Amazon for an espresso machine under $200 to see which product the AI bots choose and why.
The approach extends to social media apps. Instagram was used as an example where a bot can search for recipes and then watch and save preferred videos. It was explained that the program could have millions of bots doing this at the same time.
RELATED: China’s hottest start-up thinks it can beat the West with a new idea of freedom
There’s just one problem. The agents in the massive simulation cannot be wholly trusted to act the way humans would. As with any AI agent, there is a possibility that an agent in the world sim could fail to perform as instructed or desired.
In Harvard’s controlled study of 400 trials, measuring the bots’ adherence to 10 behavioral attributes, the results were good but not perfect. “The agents successfully expressed or correctly suppressed their assigned traits 91.5% of the time,” the researchers wrote.
MatrAIx, headed by Xiaomin Li (Harvard) and Yuexing Hao (MIT), conducted over 16,000 evaluation trials across eight different types of tasks. They used three large-language models for their studies: Claude Opus 4.8, Chat GPT 5.5, and Claude Haiku 4.5.
Human evaluations of the persona quality, however, were actually lower, listed as an average of 4.135 out of 5, which equates to about 83%.
This roughly 8.5% rate of agents simply failing to adhere to a persona could have large implications for customers who use the bots on a large scale. In the example provided for Coca-Cola consumption, that carries the possibility of 85,000 bots not answering accurately.
But the deeper problem the study confirmed is that Earth’s billions of human beings can’t be accurately simulated by AI agents programmed to be well-behaved. It’s impossible to model and predict human behavior if the agents standing in for the humans aren’t allowed to behave in all the devious, dangerous, and flawed ways individuals and groups do.
RELATED: Infamous cyberattackers claim their next shocking victim — using a decades-old trick
Mark Felix/AFP/Getty Images
That raises the dark prospect that entities or governments seeking ultimate predictive power will have to build large-scale simulators where agents representing real people are allowed to have antisocial, criminal, sinful, or even evil traits and behaviors.
The alternative, deeper human research, is nowhere near as cheap and swift as AI modeling. Most research groups seem to say that human focus groups are not effective at garnering mass public appeal anyway, while at the same time surveys can be a bit more time-consuming. The trade-off is that they are still human.
Like Blaze News? Bypass the censors, sign up for our newsletters, and get stories like this direct to your inbox. Sign up here!
Harvard, Mit, Chatbot, Artificial intelligence, Survey, Tech
