AI Simulates Live-Stream Viewers to Predict Risky Behavior

Live streaming is chaotic. Millions of people comment, react, and follow the crowd in real time. That makes it hard for platforms to predict problems like scams, harassment, or groups of bad actors forming in chat. Existing AI simulations treated users as fixed personalities based on past behavior, but that doesn't work when the stream itself changes how people act. The new system, LiveSim, was built to solve this.

LiveSim thinks of each virtual user as a "living guess." It starts with a rough idea of who they are, then watches how their simulated actions compare to real observations. When the simulation drifts from reality, the AI pinpoints what environmental factors caused the shift—maybe a sparky host, a controversial topic, or a reward bait in the chat. It then saves those lessons in a shared "collective memory," so every simulated user learns from the experiences of others, making future simulations more accurate.

The researchers tested LiveSim on real-world live-stream risk-control data. It outperformed older models at predicting individual user behavior, and it could also simulate bigger ecosystem-level trends: how risk evolves over time and how platform interventions, like banning a scammer or changing chat rules, ripple through the entire community. For platforms, this is like having a flight simulator for online safety—you can test policies without risking real users.

For everyday viewers, this means safer streaming environments. Platforms could use LiveSim to stop scams before they spread or adjust moderation policies in a virtual sandbox first. It's a promising step toward making live-streaming less reactive and more proactive—protecting users without sacrificing the lively, real-time feel that makes these platforms fun.