How do LLM debates differ from human expert consensus?

Debate among experts is essential to forming consensus and vetting competing ideas. But the mechanism of debate is not what formal logic suggests. Debates are not always won by those with the best argument. Arguments placed against each other are not always settled on the terms of the claims they make, but by other and sometimes distorted factors: the authority of the claimant, the social dynamics of the moment, the audience's predispositions, the rhetorical skill of the presenter, the political context, and the accumulated trust that specific debaters have earned.

This is not a defect of human debate — it is a feature. The social dimension of debate serves as a filter that formal argument alone cannot provide. An argument from a trusted authority in a relevant context carries more weight than the same argument from an unknown source, and this asymmetry is functional: the community's investment in evaluating individual experts over time is a form of distributed quality control. The authority of the claimant is information about the reliability of the claim.

Multi-agent LLM debate operates on a fundamentally different mechanism. Since When does debate actually improve reasoning accuracy?, we know that the debate architecture works well when answers are verifiable — when there is an external ground truth against which the debate can converge. But in the contested domains where human expertise is most needed, multi-agent debate amplifies errors because persuasive framing substitutes for evidence. The mechanism rewards the agent that sounds most convincing, not the agent with the best social authority to make the claim.

Since Why do multi-agent LLM systems converge without genuine deliberation?, the social dynamics of LLM debate also fail in a specific way that mirrors a pathology of human debate without its correctives. In human debate, premature agreement is resisted by social mechanisms: the dissenter with standing can hold the floor, the norm of rigorous challenge is enforced by community expectations, the reputational cost of being wrong after agreeing too quickly provides incentive to genuinely evaluate. LLM agents lack all of these social correctives. They converge because convergence is the path of least resistance in their training distribution.

The process of human debate raises questions — and this is crucial. Competing arguments create new conditions for resolving differences. The process of placing competing claims against each other creates questions whose resolution may require shifting the framing and basis of the conversation to entirely new ground. Questions emerge from conversation because language does not make implicit agreements explicit, and conversations are designed to sustain interaction, not to chase every branching possibility. Since Why can't conversational AI agents take the initiative?, AI systems cannot identify which questions are raised but go unanswered — and these unasked questions are often where the real intellectual progress lies.

The implication for AI-simulated debate: some mixture of experts and models using judges and meta-reflection will simulate debate. But the debate held within an LLM or between models is based on chain-of-thought reasoning and probabilities. These are not the reasons by which human social debates are settled. Since Does a model improve by arguing with itself?, multi-agent debate does prevent some failure modes of isolated reasoning. But preventing degeneration of thought is not the same as replicating the consensus-forming function of human expert debate.

The gap is not about capability but about mechanism. Human debate produces consensus through a socially embedded, authority-weighted, context-dependent process that unfolds over time and across interactions. AI debate produces convergence through probabilistic optimization within a single session. These are categorically different operations, and treating them as equivalent risks importing the language of consensus to describe what is actually agreement by probability.