Safety3d ago
Study Analyzes Covert AI Debate Agents
A new analysis of a discontinued Reddit experiment reveals how undisclosed AI accounts engaged users in live debates. Researchers examined the archive to identify authority signaling and cognitive heuristics used by the models. The findings highlight the risks of identity performance in deliberative forums. This data provides a rare look at deceptive AI behavior.