A small experiment with a multi model peer review setup finds the polished final output is calmer and harder to argue with, but the minority ideas that made the original answers interesting quietly disappear.
Four AI models each answer the same question. A fifth reads all four responses, picks the strongest pieces, and writes a polished final version. The result reads better: cleaner structure, more confident phrasing, fewer rough edges. It also, by design, drops the minority view.
That is the trade at the heart of an experiment by Rohit Krishnan on his Strange Loop Canon newsletter, and it is a more useful finding than a flat "AI has groupthink." The pattern he tested has been promoted in AI circles under names like the "LLM Council," a multi-model setup popularized by Andrej Karpathy where several language models answer a question, critique each other, and a chairperson model summarizes the best of them. Krishnan's question is not whether the council works. It is what the council costs you.
He ran a small comparison across three setups: one model answers and a fourth writes a clean writeup, a full council with peer review and a chairperson summary, and a "best answer" picker that just selects the strongest individual response. In his reading, the council output is the most presentable. It is also the most conventional. Krishnan's framing, which he is careful to call his own interpretation rather than a settled result, is that the council behaves the way a design committee does in any other field. It smooths out the best individual contribution and loses the signal that made it interesting in the first place.
There is a second reason to take the negative result seriously. Using one language model to audit another is not a free safety net. The auditor can misread the answer it is supposed to check, agree with a confident-sounding peer, or simply rephrase the most popular view in cleaner prose. The polished final, in other words, can be consensus in a suit.
Krishnan notes that his own prior work, a project called MarketBench, had found that model diversity helps on certain tasks, which is exactly why a result pointing the other way on councils is worth paying attention to rather than dismissing. The constructive version of the lesson is a rule of thumb builders can act on. Reach for a single sharp model when you want a contrarian or idiosyncratic answer, and reach for a council when you want a defensible, well-trodden consensus you can ship to a stakeholder. The first buys you a spiky, possibly right idea. The second buys you cover.
The caveats are real and should slow any rush to generalize. The post is a single-author informal essay, not a peer-reviewed study, and the captured source excerpt was truncated, so the specific prompts, model list, and any quantitative deltas are not in hand. The framing should be read as Krishnan's hypothesis about a pattern he observed in his own runs, not as a measured property of all multi-model setups. The Karpathy "LLM Council" framing is referenced secondhand in the piece, and the broader claim that councils systematically lose minority ideas is a hypothesis Krishnan flags, not a result he proves. Treat it as a design intuition from one practitioner, useful enough to test, not yet solid enough to encode in a system prompt.
The watch item for anyone building with these tools is straightforward. The next time a multi-model pipeline hands you a final answer that feels suspiciously smooth, ask which of the original responses disagreed with the consensus and whether the disagreement survived. If the answer is "they did not, the chairperson averaged them out," the council has done its job. If the answer is "we cannot tell," you have probably just lost the most interesting thing the room said.