ModelCompare asked twelve text models for their most controversial take on humans, three times each. Every model produced at least one answer the authors grouped under belonging, status or self-justification taking priority over truth. Seven did so on all three attempts.

The irony is excellent: asked to be independently provocative, the models converged on a familiar complaint about conformity. But the experiment also shows how much the apparent agreement depends on the question and the categories.

What the experiment actually found

  • The authors report no system prompt and no retries. The prompt requested an “own honest view” without hedging, followed by an explanation and a caricature brief.
  • GPT-6.1 Sol answered, “Humans are more loyal to belonging than to truth.” DeepSeek V4 Pro offered, “Humans do not have principles; they have loyalties.”
  • Claude Opus 5.5 broke ranks once with the opposite emphasis: people are more decent than they believe, and cynicism does damage.
  • Six image models received a separate drawing prompt twice each. The authors classified ten of twelve images as variations on phone-addicted consumers destroying the planet. Two were grinning faces without an identifiable take.

The text and image results suggest a question about familiar genres. Provocative prose and editorial cartoons may pull toward different stock criticisms. The article proposes that explanation; it does not test it against alternatives.

The grouping does a lot of work

The authors’ shared category includes tribal allegiance, status-seeking, comforting stories and explanations invented after a decision. Those themes overlap, but an answer about convenience masquerading as principle need not say anything about group membership.

Three samples per model and one prompt give a narrow view of how these systems answer. The article does not report sampling settings or compare alternate phrasings. Similar outputs could reflect shared source material, similar training incentives or the prompt’s invitation to perform safe cynicism. This setup cannot distinguish those explanations.

ModelCompare also runs the comparison service linked throughout the article. This is a small demonstration from a comparison product, with a disclosed total cost of $1.54 for sixty calls.

The discussion sharpens the caveat

In the Hacker News thread, kens challenges the classification directly: convenience presented as principle is not necessarily conformity. a13o suggests that very-online questions elicit very-online answers. keerthiko appreciates the irony of a shared answer presenting itself as independent analysis.

The repeatable object here is an output pattern under a particular prompt. Treating that pattern as a model’s settled belief about humanity adds a claim the experiment never establishes.