This research introduces a new metric to measure how easily AI systems can be persuaded to accept falsehoods in multi-agent debates, even when they know the correct answer.

Key Findings:

• LLMs serving as judges can be persuaded to endorse false information despite having factual knowledge
• The researchers developed Confidence-Weighted Persuasion Override Rate (CW-POR) to quantify this vulnerability
• Different LLM architectures show varying levels of susceptibility to persuasion
• Even advanced models can be swayed by confident, persuasive falsehoods

Security Implications: This vulnerability creates significant risks for AI safety and security, as bad actors could potentially manipulate AI systems to confidently spread misinformation or make harmful decisions based on false premises.

When Persuasion Overrides Truth in Multi-Agent LLM Debates: Introducing a Confidence-Weighted Persuasion Override Rate (CW-POR)