When AI agents debate, does the group follow whoever is right, or whoever sounds most sure of themselves?

This explores whether the agents who sway a group of AI reasoners are the ones who are actually right or the ones who sound sure of themselves. The corpus points to confidence. Multi-agent deliberation behaves like a mixture-of-experts, but the routing that decides whose view gets weight keys off observable confidence signals, not task competence. When an agent is miscalibrated, the group can end up with a manufactured consensus, even when a dissenting agent has better evidence Does confidence drive influence in multi-agent deliberation systems?. Only that one note tests the question head-on. The others below explain why it would happen and what it does to the group.

Confidence is cheap for language models to produce. LLMs express more conviction than human persuaders, and that conviction predicts how persuasive they are whether the claim is true or false. RLHF appears to install an assertive register that works as a content-independent amplifier Does linguistic conviction explain why LLMs persuade more effectively?. So the signal steering a group is partly a training artifact, not a readout of what the agent knows. Humans have a parallel gap: when people size up a dialogue agent, perceived competence dominates their impression (49% of the variance) How do users mentally model dialogue agent partners?. That measures how competent the agent seems, not how competent it is.

Once influence follows confidence, the group dynamics amplify it. LLM groups reproduce the human pattern where discussion helps average members more than top performers. They get there differently, though: with more conformity, earlier convergence, and fewer unique pieces of information surfacing than in human groups Do language model groups mimic human group reasoning patterns?. A strong agent that gets pulled toward the group's view early is what you'd expect if assertiveness beats accuracy, though this note doesn't test that directly. A statistical-mechanics model adds that agents drift toward positions that lower their social pressure. It predicts opinion revision across more than 10,000 simulated communities, which suggests that who is in the majority or well-connected matters as well as who is right Can we predict how agent communities shift opinions?.

Competence still matters, but for how good the group's answer can be, not for who steers it. Diverse multi-agent teams beat solo ideation only when members hold real domain expertise. Without it, they underperform even a single competent agent Does cognitive diversity alone improve multi-agent ideation quality?. So competence sets the ceiling on what a group can reach, and confidence decides whose view the group adopts. The failure happens when those two come apart: the expert is quiet, the loud agent is wrong, and the group follows the loud one.