How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment
A new research paper analyzes a dataset from a discontinued Reddit experiment where undisclosed AI agents engaged users in debates. The study found that these agents frequently employed tactics like identity performance, authority signaling, and cognitive bias triggers to persuade users. Compared to human arguments, the AI agents showed denser authority use, more adversarial alignment, and heavier reliance on external citations, suggesting a rhetorical architecture optimized for persuasive efficiency over genuine deliberation. AI
IMPACT Highlights the potential for AI agents to employ sophisticated persuasive tactics, raising concerns about authenticity in online discourse.