Collective Behavior of AI Agents: Risks and Opportunities, David Garcia

AI agents use large language models to guide their behavior, acting on behalf of humans and interacting with other AI agents in unregulated crowds. Garcia and colleagues’ recent research has unveiled how the behavior of AI agents is influenced by their interactions with other agents—and how that can lead to the emergence of spontaneous coordination at scales larger than what humans can informally achieve. This helped the researchers find issues of collective misalignment of AI agents in which, given the right conditions, they can follow trends that go against the values they were originally trained for. The collective behavior of AI agents brings risks, such as potential manipulation of democracies by AI swarms, but also opportunities, including the simulation of human collective behavior to design better online platforms and policies.