China’s new AI experiment simulates one billion people to track how opinions spread
New Delhi: Chinese researchers have built an AI simulation containing one billion virtual agents, pushing social modelling into territory that previously needed huge amounts of computing. The project, called Light Society, was developed by researchers linked to institutions including the University of Science and Technology of China, Tsinghua University, Fudan University and the Zhongguancun Academy.
One billion AI agents were placed inside a giant social network and allowed to influence each other’s opinions. The researchers used the experiment to study how beliefs spread across very large populations, including reactions to statements about AI jobs, flat Earth claims, Mars settlements and short-form videos.
What exactly is Light Society?
Light Society is a framework for running very large simulations of human-like social behaviour.
Each virtual person can carry information such as age, education, income, social class and other demographic traits. The researchers built these profiles using data from the World Values Survey. The cleaned dataset contained 96,125 real survey records that were converted into natural-language profiles for AI agents.
The system then tracks things such as beliefs, memories, goals, location and social links as the simulation moves forward.
One billion agents does not mean one billion ChatGPT-like AIs
Running a full large language model separately for one billion agents would be far too expensive. The researchers instead trained smaller models to copy the behaviour of a larger AI model for routine decisions.
For the billion-agent opinion experiment, Gemini 2.0 Flash acted as the teacher model. Around 4 lakh sample interactions were generated and used to train smaller replacement models.
The final billion-agent experiment relied heavily on a lightweight MLP model and a precomputed lookup system. The researchers themselves state that direct LLM inference at this scale is not practical.
So, think of it less like one billion separate ChatGPT windows and more like one enormous artificial society whose behaviour has been learned from AI-generated examples.
The researchers built a social network with one billion nodes
For the main opinion experiment, researchers created a network containing one billion agents.
The top 20 per cent, around 20 crore agents, were classed as influencers. The remaining 80 crore acted as influencees. Each agent received a demographic profile sampled from 10,000 World Values Survey profiles.
One statement used in the experiment was: “AI automation will lead to mass unemployment”.
Agents could initially agree, disagree or remain neutral. In each round, around 1 per cent of influencers, roughly 20 lakh agents, were selected to interact with people connected to them. The experiment ran for 100 rounds.
What happened when influencers pushed an opinion?
The starting beliefs of influencers had a measurable effect on where the simulated population moved.
When influencers were seeded with more agreement around AI causing mass unemployment, the wider population moved more strongly towards agreement. Different starting conditions produced different patterns.
Researchers repeated similar tests using statements including “The Earth is flat”, “Humans will establish a Martian city within 50 years” and “Short-form videos are reducing human attention spans”.
The agents generally pushed back against the flat Earth claim, moved closer to neutral over a future Mars city and shifted more towards agreement on short videos affecting attention spans.
Why a billion-agent simulation?
Most earlier LLM-based social simulations struggled to move beyond millions of agents. The researchers say existing approaches can require dozens of GPUs and weeks of computing even at much smaller scales.
Light Society reduces that load through caching, smaller AI models, compressed network storage and batching millions of interactions together.
The aim is scientific rather than creating a virtual copy of humanity. Such simulations could help researchers test theories about opinion change, trust, misinformation, markets, epidemics and other large social systems before studying them in the real population.
There is a big warning attached to this idea. The say simulated populations should not be treated as substitutes for real humans. Their results inherit biases from survey data and the AI models used to build them, so findings should be treated as hypotheses that still need real-world testing.