(LLMs) are used to check facts, but they can be tricked by carefully crafted inputs. Existing attacks either use meaningless character sequences that are easy to detect or rely on a single model that gets stuck in its own errors.
This paper introduces Debate-to-Deceive, a method using multiple AI agents that argue against a false claim as if in a formal debate. Each agent chooses a rhetorical style, logic (Logos), emotion (Pathos), or credibility (Ethos), and adapts its strategy through competition, producing persuasive and coherent text.
On four fact-checking benchmarks, the method outperformed existing attacks and defeated most existing defenses. This shows that current fact-checking systems are vulnerable to rhetorically styled arguments, highlighting the need for more robust defenses.