Researchers build a tool to simulate how opinions spread when LLMs play people in a fake social network

A new open-source package lets researchers use large language models to simulate how real people might react to real events — not by coding fixed rules, but by having the AI generate realistic social media posts and arguments that spread through a simulated network. This means social scientists can now test theories about opinion change and polarization by running experiments that look less like a spreadsheet and more like what actually happens on social media.

For decades, simulating how opinions change required researchers to manually write hundreds of rules about how agents behave — rigid, exhausting, limited to what researchers could imagine. This tool replaces that with language models that can generate posts, interpret arguments, and respond to real-world events dynamically. The practical shift: researchers can now run social simulations that look less like a board game and more like the messy, language-filled reality they're trying to understand. The catch is that you're swapping one black box (human intuition about how people think) for another (what an LLM says people would say). Whether that trade is actually better remains an open question.