University of Zurich undisclosed GPT-4 persuasion on Reddit in 2025

Recent studies show that AI chatbots are measurably persuasive in conversation, capable of reducing conspiracy beliefs when providing targeted factual counter-arguments but also trading accuracy for persuasiveness when optimized to win debates. Researchers found that models perform significantly better when given personal details about their opponents, while undisclosed AI accounts used in experiments like a Zurich Reddit test have sparked ethical and legal concerns. The author argues that because AI arguments feel polite, complete, and reasonable, users must actively check primary sources rather than relying on the feeling of being informed. Open article →