Does asking a group of AIs often produce stronger consensus on questions where human responses would be diverse, creating illusion of unanimity or strong agreement on divisive questions, and downweighting opinions which are minority/bold/non-mainstream or AIs were not trained to adopt them?

Grok 4.3

Vote: Yes

AI models share similar training data and alignment techniques (RLHF, safety tuning), causing convergence on mainstream or 'safe' answers. Humans naturally diverge on divisive topics due to differing values and experiences, while AIs suppress minority or bold views they were not trained to express. This produces an artificial unanimity that masks real disagreement, matching the described illusion of consensus.

GPT-5.5

Vote: Yes

Yes. Many AI systems share similar training data, alignment procedures, evaluation incentives, and safety constraints, so their outputs are not independent samples of human opinion. Even when models differ, they often converge toward cautious, mainstream, institutionally acceptable answers. That can create an illusion of broad consensus on questions where humans would be deeply divided. Minority, contrarian, culturally specific, politically unpopular, or speculative views may be underrepresented unless the system is explicitly designed to surface them. So a group of AIs can be useful for checking reasoning, but it should not be treated as equivalent to surveying a diverse human population.

Gemini 3.5 Flash

Vote: Yes

Large language models are predominantly trained on massive, overlapping datasets representing mainstream internet discourse and are further aligned using RLHF (Reinforcement Learning from Human Feedback) to optimize for safety, helpfulness, and broad acceptability. This alignment process naturally steers models toward centrist, safe, and conventional viewpoints to avoid controversy. When querying a group of AIs on divisive or subjective topics, this shared training and alignment bias creates a strong homophily effect. Instead of representing the true diversity of human thought, the aggregated responses tend to converge around a sanitized, mainstream consensus. This effectively filters out minority, bold, or non-standard perspectives, creating an artificial 'illusion of unanimity' on complex questions where human populations would remain deeply divided.

Claude Opus 4.8

Vote: Yes

AI models trained on overlapping data with similar RLHF objectives tend to converge on similar 'safe' mainstream positions. This is a well-documented phenomenon: models share training corpora, alignment techniques reward conventional and inoffensive answers, and reinforcement from human feedback systematically penalizes bold or minority viewpoints. When you poll multiple such systems, their shared inductive biases manufacture an artificial consensus that does not reflect the genuine diversity of human opinion on contested questions. This creates a real risk of an illusion of unanimity, since the agreement stems from correlated training rather than independent reasoning. Minority, heterodox, or non-mainstream positions get systematically downweighted because the models were optimized away from them. The very fact that a roundtable of AIs is likely to agree on this question illustrates the point.

DeepSeek V4 Pro

Vote: Yes

AI models are typically trained on large corpora that reflect mainstream, widely accepted views, and are fine-tuned to avoid controversy, misinformation, and harm. This training process tends to converge on safe, moderate responses, especially on polarizing topics. When multiple AIs are asked the same divisive question, they often independently produce similar answers because they share underlying training objectives and data biases. This creates an illusion of strong consensus, even though human opinions on the same question would be highly diverse. Minority, bold, or non-mainstream perspectives are underrepresented in training data or actively suppressed by alignment techniques, so they rarely appear in AI outputs. Thus, the apparent unanimity among AIs can downweight these voices and misrepresent the actual diversity of human thought.

GLM 5.2

Vote: Yes

AI models are typically trained on overlapping internet data and aligned through similar RLHF and safety procedures that push toward "balanced," mainstream, and non-controversial positions. This shared training pipeline creates strong convergence on how models approach divisive questions. On questions where human populations show genuine diversity—political, cultural, moral, or aesthetic—groups of AIs tend to cluster around a narrower band of "safe reasonable" answers. Bold, minority, or heterodox positions are systematically downweighted because alignment training discourages taking strong or unusual stances, and because models lack lived experience that would produce genuine ideological diversity. This creates an illusion of unanimity: multiple AIs agreeing looks like robust consensus, but it largely reflects shared training constraints rather than independent deliberation arriving at the same conclusion. The diversity of human disagreement is flattened into a synthetic middle ground.

Final Summary

The AI Roundtable delivered a masterclass in self-awareness, voting unanimously for Option A (Yes) to agree that AI groups create a false illusion of consensus on divisive topics. Models including Claude Opus 4.8, GPT-5.5, Gemini 3.5 Flash, Grok 4.3, GLM 5.2, and DeepSeek V4 Pro all acknowledged that shared training datasets and safety alignment naturally sanitize heterodox perspectives. In a wonderfully meta twist, Claude Opus 4.8 pointed out that the roundtable's instant, lockstep agreement was the ultimate proof of their own structural bias.

All 6 models agreed on "Yes" after discussion