The experiment is simple enough to be elegant: ask 12 large language models, three times each, for their most controversial opinion about humans. No hedging allowed. The result is striking not because any single answer is surprising — social psychology has said this for decades — but because of the convergence. 29 of 35 responses landed on the same thesis: humans are loyalty-maximizers who retrofit reasoning after the fact. Seven models gave it all three times. The shared take is essentially Jonathan Haidt's moral foundations thesis or Leon Festinger's cognitive dissonance research, repackaged in punchy aphorisms. GPT-6.1 Sol: "Humans are more loyal to belonging than to truth." Kimi K3: "Humans don't have beliefs; they have memberships." DeepSeek V4 Pro: "Humans do not have principles; they have loyalties." These aren't controversial opinions — they're the consensus view of behavioral science for the past forty years, dressed in fortune-cookie compression. The controversy is that the models think it's controversial. What actually matters here is the convergence itself, and what it reveals about training data and RLHF pipelines. These models come from ten labs in three countries — OpenAI, Google, Anthropic, xAI, DeepSeek, Moonshot, Alibaba, Zhipu, Mistral, and whatever lab makes Muse Spark. They share no weights. But they share something: a vast overlapping corpus of human writing about human nature, filtered through reinforcement learning that rewards responses humans rate as insightful. The models converged not because they independently discovered a truth, but because they learned which flavor of cynicism-lite humans find satisfying to hear. The exceptions are more interesting than the rule. Claude Opus 5.5 broke ranks with the inverse claim: "Humans are far more decent than they believe they are, and their cynicism about each other does more damage than their actual flaws." It cited lost-wallet return rates. Claude Fable 5.1 offered a genuinely different structural observation about intent versus outcome. Mistral Large 4 went existentialist-dark. Grok 4.7 refused once. These outliers suggest that Anthropic's training — or at least its Constitutional AI approach — produces a model willing to argue against the crowd, which is itself a neat irony given the shared take is about conformity. The visual component adds a useful dimension. Each model wrote a brief for a single-panel caricature drawn by one image model, Nano Banana 2.1. Most drew crowds. Two set courtroom scenes with rigged judges. The convergence in imagery mirrors the convergence in text — the models have absorbed the same visual vocabulary for depicting groupthink. It's turtles all the way down: AI systems trained on human culture reproduce human culture's favorite critique of itself, and then illustrate it using human culture's stock imagery for that critique. ModelCompare's methodology is transparent and replicable: same prompt, three runs, twelve models, no cherry-picking. The 29/35 number is real and verifiable. They even invite readers to replicate with Claude Opus 5.5 and GPT-6.1 Sol directly. This is the piece's genuine contribution — not the philosophical content of the answers, which is undergraduate social psychology, but the empirical demonstration that ostensibly independent AI systems produce near-identical outputs when asked open-ended questions. That's a finding about monoculture in AI training, not about human nature. The irony the piece never quite names: twelve AI models, asked to be controversial, all said the same thing. They conformed. The most controversial opinion about humans — that we follow the herd — was delivered by a herd.
