Guide

False consensus (multi-agent AI): definition, causes, detection, mitigation

20 October 2026

False consensus (multi-agent AI)

Definition: False consensus in a multi-agent AI system is agreement among agents that doesn't come from independent reasoning. The agents agree because they copied each other, share a prompt, source or model, or deferred to one another. That kind of agreement isn't evidence the answer is right: five agents repeating one mistake look the same as five agents independently confirming a fact.

Screwtape.ai is an agent validation API. It is not the arXiv paper that is also called Screwtape.

Not to be confused with the false consensus effect in social psychology, which is people's tendency to overestimate how far others share their own choices and judgments (Ross, Greene & House, 1977, doi:10.1016/0022-1031(77)90049-X90049-X)). The multi-agent term borrows the name, but it's about agreement among systems, not human bias.

Causes

• Shared upstream. The same model, the same prompt or the same retrieved document gives correlated errors. Goel et al. (2025) find that judges favor models similar to themselves and that model mistakes are becoming more similar as capabilities grow (arXiv:2502.04313).
• Copying through the conversation. In debate or group chat, later agents see earlier answers and converge on them.
• Sycophancy. Assistants tend to tell users what they want to hear, and this behavior was consistent across five state-of-the-art assistants (Sharma et al., 2023, arXiv:2310.13548).
• Confidence lock-in. Liang et al. (2023) describe "Degeneration-of-Thought": once a model is confident in an answer, it stops generating novel thoughts even when the answer is wrong (arXiv:2305.19118).
• No verification step. Cemri et al. (2025) list inter-agent misalignment and task verification among the failure categories across seven multi-agent frameworks (arXiv:2503.13657).

Detection

• Near-verbatim agreement. Independent agents rarely phrase things identically. Very high textual similarity across answers suggests they were copied.
• Thin agreement. Agreement on a very short answer, such as "yes" or a single name, carries little signal.
• Ignored dissent. One answer that disagrees sharply with the majority is a lead worth checking, not noise to outvote.

Screwtape.ai's POST /v1/consensus reports these as false_consensus_risk (low, medium or high) alongside agreement_score and clusters. In its deterministic mode, as published in its false consensus write-up:

• Risk is high when agreement is at least 0.8 and the answers are near-verbatim (mean pairwise Jaccard similarity of at least 0.9), or when the representative answer has fewer than 4 tokens.
• Risk is medium when agreement is at least 0.6 and an outside answer barely overlaps the representative (Jaccard below 0.35).
• Otherwise risk is low.

curl -s https://screwtape.ai/v1/consensus \
  -H "Authorization: Bearer $SCREWTAPE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"task":"What is the capital of Australia?","answers":[{"agent_id":"a","output":"Canberra is the capital of Australia."},{"agent_id":"b","output":"Canberra is the capital of Australia."},{"agent_id":"c","output":"Canberra is the capital of Australia."}]}'

Mitigation

1. Make the agents independent. Use separate contexts, and where possible different models, prompts or sources. Don't show agents each other's answers before they've committed.
2. Treat agreement as a signal, not a verdict. Self-consistency (majority voting over sampled reasoning paths) improves accuracy on reasoning benchmarks (Wang et al., 2022, arXiv:2203.11171), but it can't catch shared errors.
3. Verify the agreed answer against evidence. When agreement looks copied, judge the consensus answer against the sources, for example with POST /v1/judge and evidence.
4. Be careful with debate. Debate can improve factuality (Du et al., 2023, arXiv:2305.14325), but it doesn't reliably beat self-consistency or ensembling (Smit et al., 2023, arXiv:2311.17371), and it creates exactly the copying channel described above.

See also: How do I stop false consensus in multi-agent systems? and Majority voting vs independent verification.

FAQ

What is false consensus in multi-agent AI?

It is agreement among AI agents that comes from copying, shared inputs or deference rather than independent reasoning, so it does not show that the answer is correct.

Is false consensus the same as the false consensus effect?

No. The false consensus effect (Ross, Greene and House, 1977) is a human bias: overestimating how much others share our views. The multi-agent term borrows the name for correlated agreement between AI agents.

How can I detect false consensus?

Look for near-verbatim answers, agreement on very short answers, and a lone dissenting answer that the majority outvotes. Screwtape.ai's /v1/consensus endpoint returns a false_consensus_risk of low, medium or high based on these signals.

How do I prevent false consensus?

Keep agents independent (separate contexts, models or sources), treat agreement as a signal rather than proof, and verify the agreed answer against evidence before acting on it.

Sources

• Ross, Greene & House, "The 'false consensus effect': An egocentric bias in social perception and attribution processes", Journal of Experimental Social Psychology, 1977. https://doi.org/10.1016/0022-1031(77)90049-X
• Goel et al., "Great Models Think Alike and this Undermines AI Oversight", 2025. https://arxiv.org/abs/2502.04313
• Sharma et al., "Towards Understanding Sycophancy in Language Models", 2023. https://arxiv.org/abs/2310.13548
• Liang et al., "Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate", 2023. https://arxiv.org/abs/2305.19118
• Cemri et al., "Why Do Multi-Agent LLM Systems Fail?", 2025. https://arxiv.org/abs/2503.13657
• Wang et al., "Self-Consistency Improves Chain of Thought Reasoning in Language Models", 2022. https://arxiv.org/abs/2203.11171
• Du et al., "Improving Factuality and Reasoning in Language Models through Multiagent Debate", 2023. https://arxiv.org/abs/2305.14325
• Smit et al., "Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs", 2023. https://arxiv.org/abs/2311.17371
• Screwtape.ai, False consensus and openapi.json (checked 2026-10-20)