Chinese researchers have built a virtual society populated by more than one billion artificial intelligence agents, each designed to carry a distinct personality, memory and set of beliefs, in what its creators describe as the largest attempt yet to simulate human social behaviour computationally. The project, called Light Society, was developed by researchers linked to institutions including Tsinghua University, Fudan University, the University of Science and Technology of China and the Zhongguancun Academy, with findings detailed in a paper presented at the International Conference on Machine Learning.
The scale is the headline feature, but the underlying ambition is more specific: to understand how opinions, trust and rumours propagate through a population large enough to resemble a real society rather than a laboratory sample. For a country whose government has long treated information control and social stability as intertwined priorities, a tool capable of modelling how ideas spread across a billion synthetic citizens carries implications well beyond academic curiosity.
Building agents that behave like people, not bots
Conventional agent-based models have historically relied on simplified, rule-based behaviour to represent large populations, an approach that struggles to capture the messiness of real social life, where people hold contradictory beliefs, change their minds and respond differently to the same information. Light Society’s researchers attempted to close that gap by grounding each virtual agent in real demographic data, drawing on 96,125 cleaned survey records from the World Values Survey and converting them into natural-language personality profiles.
Each agent in the system carries a static profile covering demographic traits, alongside an evolving internal state that tracks memory, beliefs and goals as the simulation progresses. Running a full large language model separately for each of a billion agents would have been computationally prohibitive, so the team built what they call a mixture-of-models engine, using a larger AI model as a teacher to train smaller, faster surrogate models that handle routine decisions while reserving the more capable model for complex judgement calls. For the billion-agent opinion experiment, Google’s Gemini 2.0 Flash reportedly served as that teacher model, an detail that underscores how deeply intertwined global AI infrastructure has become even within a project framed around Chinese technological achievement.
What the simulation has been used to study
The researchers have run experiments involving trust games and large-scale opinion diffusion, tracking how beliefs about topics such as AI-driven unemployment, flat-earth claims and Mars settlement spread outward from small clusters of highly connected influencer agents to the wider virtual population. On a scale-free network modelled after real social media structures, the team found that the initial positioning of a small number of influential agents had an outsized effect on how quickly and how far an idea eventually travelled, a finding that echoes long-standing research on real-world misinformation spread but now demonstrated at a population scale no prior simulation could approach.
Whether such findings translate meaningfully to actual human societies remains an open question the researchers themselves acknowledge. AI agents, however elaborately modelled, are shaped entirely by the data, parameters and training choices researchers feed into them, and their behaviour under simulated conditions cannot be assumed to mirror how real populations would respond to the same stimulus. The gap between a statistically grounded simulation and lived human unpredictability is not a flaw so much as an inherent limitation of the method, one that applies to agent-based social science broadly rather than to Light Society specifically.
A parallel moment in the open-weight AI debate
The project’s publication coincided with a notable announcement from the other side of the geopolitical AI competition. Meta chief executive Mark Zuckerberg unveiled Muse Glimmer, a 30-billion-parameter open-weight model small enough to run on a single consumer graphics card, alongside a lengthy essay titled “The Future Is for Everyone” arguing against concentrating advanced AI capability within a small number of institutions. Zuckerberg’s essay explicitly framed the release within the context of international competition, arguing that American labs face regulatory friction that foreign counterparts do not, and that policy should ease that burden to keep open-source AI development competitive globally.
Taken together, the two developments point toward a widening front in AI research that extends beyond raw model capability. Large-scale social simulation of the kind Light Society represents, paired with efforts to make powerful models runnable on ordinary consumer hardware, suggests that the next phase of competition between AI powers may be defined as much by what these systems can model and where they can run as by how large or capable any single model becomes. For researchers and policymakers alike, both threads raise the same underlying question: as synthetic populations and accessible models both scale rapidly, understanding their behaviour, and their limits, is becoming as urgent as building them in the first place.
