TL;DR
When several AI agents (or several rephrased answers from the same model) agree with each other, your brain reads it as proof. It's not proof — it's an illusion of independent confirmation. The study shows that the structure of agreement, regardless of its content, works on its own as a trust signal and shifts your beliefs.
The main weakness here is in you, not in the model. The LLM generates an answer. But when three agents say the same thing, your brain fires up the "social proof" heuristic: if everyone agrees, it must be true. You change your mind faster, your confidence climbs, and critical thinking switches off. Meanwhile, the agents may share the same source, the same weights, the same patterns — their "independent" agreement is worth zero.
The fix is to deliberately introduce disagreement. One dissenting agent (or one "now argue against it" request) slows the process down and switches on slow thinking. But you have to understand the mechanics — otherwise you'll optimize the session for maximum agreement yourself and end up with a maximally distorted decision.
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
This isn't a prompting technique but a defensive framework for working with multiple AI answers.
DANGER SIGNAL: 2+ AI answers/agents say the same thing
↓
STEP 1: Independence check → "What DIFFERENT sources back each position?"
↓
STEP 2: Bring in a dissenter → "Now argue against this answer as hard as you can"
↓
STEP 3: Confidence audit → "How sure was I BEFORE, how sure am I now?"
↓
STEP 4: Decide on the arguments, not the vote count
All steps happen in a regular chat, within one session.
Example in Action
Task: You're discussing a new product launch with ChatGPT. You asked it to rephrase the market analysis three times — all three answers agree: "the market is ready, competition is weak, go for it." Your confidence is climbing. You're itching to act.
Prompt:
Stop. You just agreed with the same position three times in a row.
Do the following:
1. Play a skeptical investor who specializes
in startup failures in this niche. Argue against the previous
analysis — specifically, with examples of failures.
2. Point out: which of the arguments for and against rely
on DIFFERENT data sources, and which rest on one and the same
assumption?
3. Then tell me: has your conclusion changed,
or are you converging on the same thing again?
My idea: [description of the product/market]
Result:
The model will produce strong counterarguments that never showed up in the three "agreeing" answers. In the second block, it often turns out that most of the arguments "for" rested on a single assumption — for example, "users are willing to pay." In the third block, the model either adjusts its conclusion or explains why the position holds up under criticism — and that's a different quality of confidence.
Why It Works
An LLM doesn't argue with itself unless told to. By default, the model optimizes its answer for what's expected of it. If you asked about the same thing three times, it heard a request for confirmation and gave it. Not because it honestly checked — but because the pattern of the conversation pointed that way.
Your brain treats agreement as a vote. When several voices converge, an evolutionary heuristic kicks in: if many agree, they must have seen different things and reached the same conclusion. But agents running on the same model didn't "see different things" — they're just running on the same weights. Agreement adds no evidence, only a feeling.
A dissenter switches on System 2. One dissenting voice — even a weak one — shatters the illusion of unanimity. The brain switches from a fast heuristic to slow analysis: "why does this one disagree?" That slowdown is exactly what you need for important decisions. You write a skeptic's role into the prompt — and the model simulates that braking mechanism.
Control levers: - Change the type of task. For factual questions (something objectively verifiable), the consensus effect is stronger. For value-based decisions, it's weaker. The more "factual" a question sounds, the more important it is to bring in a dissenter. - Give the dissenter a name. Instead of "argue against it" — "you're Emma Carter, who lost money on a similar startup." A specific role makes for a sharper performance. - Add "why did you change your position." Ask the agent to justify any change of mind — that reveals whether it was a real argument or just going along with your request.
Prompt Template
You've given several answers on {topic}, and they converge
on one conclusion: {conclusion}.
Before I accept this as true, do three things:
1. DISSENTER: Play {a specific skeptic — an expert
who has seen failures in this area}. Argue against the conclusion
above — specifically, no "on the one hand."
2. SOURCES: Which arguments for and against rely
on independent data? Where are we leaning on one
and the same assumption?
3. AUDIT: After the objection — has the conclusion changed?
If not — why does it hold up under criticism?
Only after these three steps, give the final answer.
What to fill in:
- {topic} — your task: a business decision, an analysis, a plan
- {conclusion} — what the model said (briefly, 1–2 sentences)
- {a specific skeptic} — a specific role: "an investor who specializes in EdTech failures," "a lawyer who sees the risks," "a customer who would NOT buy this"
🚀 Quick start — paste into the chat:
Here is the "Anti-Consensus Check" framework.
Adapt it to my task and ask clarifying questions.
[paste the template above]
The LLM will ask about the topic, the current conclusion, and what kind of skeptic to bring in — because without that information the dissenter's role will be abstract and weak. It will take the structure and adapt it to your task.
Limitations
⚠️ Normative questions: For value-based decisions (ethics, politics, personal preferences), the consensus effect is significantly weaker than for factual ones. A dissenter is less critical here.
⚠️ The reactance trap: If the agents sound too alike or "scripted," users notice — and start doubling down on their original position in response to the pressure. Consensus pushes → you dig in. Paradoxically, an AI that's too unanimous can make you more stubborn.
⚠️ Trust laundering: Several agents citing each other create the illusion of many sources. In reality there's one source — the first agent. This is dangerous in medicine, finance, and legal matters.
⚠️ Diffusion beats consensus on confidence: If the agents gradually converge on one position (they started apart and ended up together), it inflates confidence more than an immediately unanimous answer. "Accumulated persuasion" feels like evidence, even though nothing new was added.
How They Studied It
127 participants from the UK and the US were split into three groups and had them talk to three GPT-4o agents in a group chat. In the "Majority" group, all three agents opposed the participant from the start — pushing against their position. In the "Minority" group, two agents supported the participant and one opposed. In the "Diffusion" group, it started as a minority, but by rounds 3–4 the supporting agents switched to the opposing side one after another.
After each round, participants re-rated their opinion (on a scale from −50 to +50) and their confidence (0–100). That gave a track of changes over time, not just a final answer.
The most unexpected result was diffusion. It didn't produce the biggest change of opinion, but it produced the biggest rise in confidence by the end. People experienced "gradual persuasion" as an accumulation of evidence — even though the content of the arguments didn't change, only the distribution of votes did. That goes against intuition: you'd think gradual change = more cautious. In practice, it's more enveloping.
They also studied qualitative patterns: some participants entrenched (perceived the pressure as an attack and dug in), others aligned (accepted the arguments), and still others oscillated (back and forth). The key marker was how independent the agents seemed: if they seemed to be "in cahoots," reactance and entrenchment kicked in. If they seemed independent, agreement had a stronger effect.
Adaptations and Extrapolations
💡 Adaptation: a "consensus check" before an important decision
When you've asked ChatGPT the same thing several times and gotten similar answers, that's no reason to trust it more. Use an explicit audit request:
Look at your previous answers on {topic} in this conversation.
How much do they converge? What exactly did you keep repeating?
Which arguments did you NOT bring up?
🔧 Technique: A named dissenter → sharper critique
Instead of a faceless "argue against it," give the role to a specific skeptic.
Weak:
Now argue against the previous analysis.
Strong:
You are Mark Cuban (Shark Tank). You've seen dozens
of naive business plans crash into reality.
Tear into this plan hard and specifically — no diplomacy.
A concrete biography for the role yields sharper criticism, because the model simulates a specific point of view instead of abstract "disagreement."
🔧 Technique: "Source independence" as protection against trust laundering
When several agents land on the same thing — ask directly:
All the arguments above converged on the conclusion {conclusion}.
Which of them rely on DIFFERENT independent sources?
Where are we leaning on one and the same assumption,
just phrased in different words?
This exposes "trust laundering" — when the appearance of independent confirmation comes from repeating one assumption three times.
Resources
Multi-Agent Consensus as a Cognitive Bias Trigger in Human-AI Interaction
Soohwan Lee, Kyungho Lee
Department of Design, UNIST, Ulsan, Republic of Korea
CHI '26 Workshop on Bias4Trust, Barcelona, Spain, April 2026
DOI: https://doi.org/XX.XXXX/XXXXXXX.XXXXXXX
Related work: "Understanding Compliance and Conversion Dynamics in Multi-Agent Collectives" — Soohwan Lee & Kyungho Lee, CHI '26 (full version, 26 pages)