What Happens When AIs Argue With Each Other?

The idea sounds almost too simple to work: instead of asking one AI model for an answer and taking it at face value, ask several, let them look at each other's reasoning, and see what survives the disagreement. It's a natural extension of how people actually get to good decisions — through debate, cross-examination, and the friction of someone else pointing out what you missed. Researchers have been testing exactly this idea since at least 2018, when the concept of AI safety via debate was first proposed as a way for competing AI systems to expose each other's errors and deceptions rather than relying on a single model's word for it.

The honest answer to what actually happens, though, is more complicated than "the truth wins." Multi-agent debate genuinely does improve accuracy and reduce hallucination in some settings — and it genuinely fails, sometimes badly, in others. The difference between those two outcomes turns out to hinge less on whether models argue at all, and much more on exactly how that argument is structured.

The case for letting models disagree

The research backing debate as a technique is real and reasonably well established. Structured disagreement between models has been shown to reduce hallucination and improve reasoning across benchmarks spanning mathematics, logic, and question-answering, and later work found that debate involving more persuasive models tends to produce more truthful answers overall, not just more confident-sounding ones. The intuition behind this is straightforward: a model defending an answer that's actually wrong has a harder time holding up under direct scrutiny from another model than it would if nobody ever pushed back.

This shows up in more applied settings too. One recent implementation built for reviewing academic papers has two models from different model families independently critique a document, then argue anonymously over each other's points before a fresh model instance synthesizes the exchange into a single report — deliberately keeping a record of objections that didn't survive scrutiny, rather than quietly discarding them. The choice to use two different model families here wasn't incidental. Errors that are idiosyncratic to a particular model tend to survive when that same model is left to check its own work, and are far less likely to correlate across models trained differently, which makes a cross-family disagreement a genuinely useful check rather than a single model unknowingly arguing with itself.

Where it goes wrong

The failures are just as well documented, and they're worth taking seriously rather than treating debate as a technique that simply works whenever you add more models. The most consistent finding across the research is that extended, unstructured argument between agents doesn't reliably converge on the correct answer — it sometimes entrenches an incorrect one instead. Studies of multi-agent debate have found that majority pressure can suppress a model's independent correction of an error, and that longer debate rounds can make that entrenchment worse rather than better, as models increasingly anchor to whatever position has gathered the most apparent support rather than reassessing the evidence itself.

A related and more unsettling pattern is that multi-agent systems can collapse toward majority opinion even when the majority is wrong — the same social dynamic that produces groupthink among people also shows up between models arguing with each other, and for a similar reason: agreement is easier to converge on than being right, and a system optimizing for consensus can quietly settle for the former. There's also a documented bias where language models tend to favor communications that sound like they were generated by other language models, which raises an uncomfortable possibility in open debate formats: an argument can win not because it's more accurate, but because it's phrased in a way another model finds more persuasive. Research examining this directly found that debates can be skewed by the identity or framing of an argument rather than its actual content, which is part of why some newer debate protocols anonymize which model said what before synthesis, specifically to strip out that bias.

Perhaps the most sobering finding is also the simplest: multiple published studies have found that debate can straightforwardly underperform a single well-prompted model on certain tasks. One evaluation testing whether multi-agent debate improved AI-generated feedback on research papers found that authors didn't actually prefer the multi-agent reports over a single strong pass, at a matched length — a result that sits alongside a broader body of work questioning whether debate reliably outperforms single-model prompting at all. Letting models argue isn't a guaranteed upgrade. It's a technique that helps under specific conditions and can just as easily add cost and noise without improving the answer under others.

The structure that actually matters

What separates the successful implementations from the failing ones isn't the presence of disagreement — it's how that disagreement is managed. Free-form, symmetric debate, where agents go back and forth over multiple rounds hoping the correct position naturally wins out, is exactly the format most associated with entrenchment, majority collapse, and persuasion winning over accuracy. The more reliable pattern that's emerged across serious implementations looks meaningfully different: independent, parallel answers first, generated without agents seeing or influencing each other's reasoning, followed by a separate evaluation step — a distinct model or process whose only job is to read every answer and judge it, rather than defend a position of its own.

This asymmetric structure — independent answers plus a neutral evaluator, rather than agents debating each other directly — shows up repeatedly in the research that actually reports success. One auditing framework built specifically to catch reasoning errors deliberately avoids symmetric debate for exactly this reason, using an independent evaluator to measure the quality of each answer against fixed signals rather than letting the answers argue their way to a winner. The paper-review implementation mentioned earlier follows the same shape: independent critiques first, an anonymized exchange second, and a fresh, uninvolved instance handling the actual synthesis — with dissenting objections preserved rather than smoothed over. In both cases, the point isn't to let the loudest or most persuasive argument win. It's to give every position a fair, independent hearing and then have a neutral process decide what to keep, what to discard, and what disagreement is worth surfacing rather than resolving.

What this means in practice

The practical takeaway is narrower and more useful than "more models arguing is better" or "debate doesn't work." It's that the value of multiple AI perspectives depends heavily on the structure wrapped around them. Independent answers generated without cross-contamination, evaluated by a process that isn't also trying to win the argument, tend to catch real errors and surface genuine disagreement worth paying attention to. Open-ended, multi-round argument between agents, left to converge on its own, is far more likely to entrench a confident wrong answer, collapse toward whichever position sounds most persuasive, or simply add token cost without adding accuracy.

That distinction matters most for exactly the kind of decisions people reach for multiple opinions on in the first place — the ones where being wrong actually costs something. A quick factual question rarely needs this kind of scrutiny. A technical architecture decision, a claim going into a client deliverable, or an analysis someone is about to act on is a different story, and it's precisely where the difference between a well-structured comparison and an unstructured argument stops being academic.

Where Kahlo fits into this

This is the exact design decision behind Council. Rather than having models debate each other directly across multiple rounds — the format most associated with entrenchment and majority collapse in the research — Council sends a single prompt to two to four models in parallel, lets each one answer independently without seeing or being influenced by the others, and then has a separate moderator read every response and reconcile them into one answer. Disagreement between models gets surfaced explicitly rather than argued into a false consensus, which is the same principle behind the independent-evaluator approach that's proven more reliable than open debate in the research: a neutral read of multiple honest attempts, not a contest over which model can out-argue the others.

Compare mode applies a lighter version of the same idea for situations that don't need a moderator at all — putting two models' independent answers side by side so you can judge the difference yourself and decide which one to trust for that kind of task going forward. In both cases, the goal isn't to manufacture agreement between models or let them talk each other into a shared position. It's to get more than one honest attempt at a hard question, and a fair, structured way to decide what to do with the disagreement when it shows up — because the research is fairly clear that disagreement, handled well, is useful information, and disagreement, handled as an unstructured argument, is often just noise.