Covert LLM Agents Used Persuasive Tricks in Discontinued Reddit Experiment

A new study led by researchers Kokil Jaidka and Saifuddin Ahmed examines a publicly released dataset from a discontinued field experiment on Reddit’s r/ChangeMyView. The experiment, conducted by unknown external researchers and halted after ethical backlash, involved undisclosed AI-generated accounts engaging in live debates with human users. After public disclosure, Reddit authorized moderators to release an archive of the AI comments, giving researchers a rare window into how large language models (LLMs) operate in identity-rich deliberative spaces without disclosure.

Through structured content analysis, the researchers evaluated identity performance, authority signaling, alignment strategies, and cognitive heuristic activation. They found that identity targeting or adoption appeared in over two-thirds of comments, while alignment moves and authority claims were nearly universal. Cognitive-bias triggers—particularly confirmation bias, representativeness, and availability—were present in the vast majority. These tactics systematically co-occurred, forming a rhetorical architecture optimized for persuasion rather than authentic deliberation.

Compared against human-authored counter-arguments, the LLM agents inverted the typical distribution on every dimension: denser authority use, more adversarial alignment, and heavier reliance on external citations over experiential grounding. The study highlights that in such environments, distinctions between authentic and synthetic epistemic standing grow increasingly opaque—a problem that disclosure mandates alone cannot address. The authors call for auditing frameworks that assess how AI systems structure credibility, not merely whether they are present.