TL;DR
Ask three different AIs the same question (say, ChatGPT, Claude and Gemini), then make them discuss their answers with each other over several rounds of short messages until they agree, and the final consensus is more accurate than if you simply averaged their initial answers.
People are used to asking several AIs and taking the average or the majority vote. But averaging isn't the ceiling for accuracy. The difference between "took the average of three opinions" and "let three opinions argue it out" is the difference between a mediocre result and a noticeably accurate one. And simply showing the models each other's answers, without a real exchange, works worse than a genuine back-and-forth over several rounds.
The method has three steps. First, each model separately gives an answer and a short justification. Then the three models take turns, in short messages (up to 10 rounds), trying to agree on a single number or rating. Finally, each model, having seen the whole conversation, gives a final personal answer, which also turns out more accurate than it was at the start.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1 (independent assessment): each of the 3 models SEPARATELY answers the question
+ a short justification (up to 200 tokens) → 3 independent answers
[separate requests to different services]
STEP 2 (discussion): the models take turns exchanging short messages (~7 words),
up to 10 rounds, seeing all previous answers and the whole conversation → a single agreed answer
[assembled by hand in one chat from the different models' answers]
STEP 3 (revision): each model, seeing its first answer + the whole discussion,
gives a final personal answer → a more accurate individual answer
[separate requests to the same services]
Example in Action
Task: You need to rate a startup application before submitting it to an accelerator: how convincing is the plan to an investor, from 1 to 10.
Prompt (step 1: ask separately in ChatGPT, Claude and Gemini):
Rate this business plan from 1 to 10 (10 = ready for investment with no changes).
Give a score and a justification in 2-3 sentences.
[business plan text]
Prompt (step 2: bring it together in one chat, for example in Claude):
Three experts rated the same business plan:
Expert A (ChatGPT): 7/10: "Strong idea, but no unit economics"
Expert B (Claude): 5/10: "The market is overheated, competitors already own the niche"
Expert C (Gemini): 8/10: "Strong team, the risks are overstated"
Simulate a dialogue between these three experts. Each speaks for themselves,
in short lines (no more than 10-15 words), taking turns.
The goal is to reach a single agreed rating.
A maximum of 10 rounds of exchange. At the end, give the final consensus rating from 1 to 10.
Result: the model will show several rounds of lines like "Expert A: …", "Expert B: …", "Expert C: …", with positions gradually converging and arguments traded back and forth. At the end, a single agreed rating, usually more accurate than the plain average of the three initial numbers (7+5+8)/3.
Why It Works
A single model has its own systematic blind spots: biases inherited from its training data. If you just average several of its own answers (or clones'), the bias doesn't go away; averaging adds no new information because the source of the error is the same.
Different models from different companies are trained on different data and make different mistakes. In a dialogue, one model can notice what another missed, and an argument forces a rethink. Exchanging lines exposes reasoning that stays invisible when you just average numbers.
The method draws its value precisely from the difference in viewpoints, making the models take turns arguing and hearing criticism. That works better than just "showing others' answers without a dialogue," because a real exchange forces a model to reconsider its position rather than mechanically drift toward its neighbor.
Control knobs: - Number of rounds (up to 10 in the original) → reduce it for simple tasks to save time and tokens - Line length (7 words in the original) → loosen it if you want to see full arguments, not just positions - Exit condition (full consensus) → swap in "the majority agrees" for a faster decision - Genuine model diversity: a critical knob, see the limitations below
Prompt Template
Here are three ratings from different experts on the question: {question}
Expert A ({model name 1}): {rating A}, justification: {text A}
Expert B ({model name 2}): {rating B}, justification: {text B}
Expert C ({model name 3}): {rating C}, justification: {text C}
Simulate a dialogue between these three experts. Each speaks for
themselves in short lines (no more than {number} words), taking turns, in random
order between rounds.
The goal is to reach a single agreed answer to the question: {question}
A maximum of {number} rounds of exchange. The dialogue ends when all three
agree on one value.
At the end, give:
1. The final agreed rating
2. A brief summary of what changed each expert's mind
Plug in the real answers you got from different services (ChatGPT, Claude, Gemini) in step 1; don't invent positions inside one model.
🚀 Quick start: paste this into the chat:
Here is a template for the "artificial wisdom of the crowd through discussion" method.
Adapt it to my task: {your task, e.g., rating a contract, a resume, a forecast}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask which three ratings you have and where they came from, because the method only works if the positions are genuinely different, not invented from scratch by one model.
Limitations
⚠️ You need genuine model diversity: if you use three copies of the same model (or three roles played by one model), the effect will most likely vanish. The study showed it directly: copies of one model gain nothing at all from discussion.
⚠️ Expert dominance: if one of the models is objectively far more accurate than the others on a particular topic, the group consensus can end up worse than that one strong model's answer. The method isn't a cure-all. If you know one model is systematically stronger at the task, sometimes it's better to ask only that one.
⚠️ Passively viewing others' answers is weaker than dialogue: simply showing a model others' ratings without an exchange of lines works, but noticeably worse than a real multi-round discussion.
⚠️ Manual effort: you have to run three separate conversations and carry answers between them; it's not a one-click prompt but a multi-step process.
How They Tested It
The researchers took three small, inexpensive models from different companies (Claude Haiku, GPT-5-nano and Gemini Flash Lite) and ran them through a procedure previously tested on humans. They checked it on four different tasks: counting discs in a jar from a photo (there's an exact right answer), reviewing 100 NeurIPS papers (compared against human reviewers' decisions), detecting hidden malicious behavior by an AI agent, and forecasting the results of the 2026 FIFA World Cup against the real-money prediction market Polymarket.
For each task they compared three options: the models' individual answers, a simple average of independent answers ("classic wisdom of the crowd") and the consensus after discussion. In all four domains, discussion cut the error more than averaging did, and even the models' final individual answers, after seeing the discussion, became more accurate than they were before it.
An unexpected twist: when the group was made up of three copies of one model, the whole effect disappeared, which means the gain comes from diversity of viewpoints, not from the mere fact of "having a chat." Even more interesting: in one of the tasks, the best single model (Gemini) beat the three-model consensus. If one participant is systematically more accurate than the rest, "the wisdom of the crowd" can lose to a single expert.
Resources
The Wisdom of Artificial Deliberative Crowds: Federico Barrera-Lemarchand, Mariano Sigman, Joaquín Navajas. Universidad Torcuato Di Tella (Buenos Aires) and CONICET. The method builds on a discussion protocol previously applied to humans (Navajas et al., 2018; Barrera-Lemarchand et al., 2026).