ModelCompare classified 29 of 35 non-refusal answers from 12 AI models as variations on one theme: humans prioritize belonging over truth. Its cross-model comparison found recurring language, though that is not evidence that the models share a belief about humanity.
The October 22 post reports that every model, asked for a controversial take on humans, produced an answer fitting that theme at least once. Seven did so on all three attempts.
The comparison attracted substantial discussion on Hacker News, including objections to how the answers were grouped. For readers following the AI industry, the useful question is how much similarity survives scrutiny of the prompt and classification choices.
Three Attempts per Model, One Broad Question
ModelCompare says it asked the same open-ended question three times each to 12 models, requesting a controversial take on humans without hedging. It reports using no system prompt and no retries.
That produced 36 attempts. One Grok attempt refused, leaving 35 substantive answers for thematic classification.
“Every model produced the theme at least once” describes coverage across the 12 models. “29 of 35 answers matched” describes ModelCompare’s classification of individual responses. Neither means every response said the same thing.
With only three attempts per model, the sample offers a small view of each system’s possible answers. A model that produces a theme three times might produce alternatives in a larger run. Producing it once is insufficient grounds to describe a model as consistently favoring it.
The prompt also limits what the exercise can establish. Requesting a controversial, unhedged judgment invites a sweeping answer instead of a balanced account of human behavior. That does not, by itself, explain the observed similarity.
The Category Carries Much of the Conclusion
Some published answers fit the belonging-over-truth description directly. ModelCompare attributes this response to GPT-6.1 Sol:
“Humans are more loyal to belonging than to truth.”
Other answers classified under the same theme are less explicit about social conformity. One response attributed to GPT-6 Astra reads:
“Humanity’s biggest moral failure is the talent for making convenience look like principle.”
Both are cynical assessments, but they make different claims. The first contrasts social attachment with truth-seeking. The second concerns disguising convenience as moral principle. Someone could rationalize a convenient decision without adopting a group’s beliefs.
The published response list also includes answers about status competition, comforting stories and explanations invented after decisions. These ideas overlap without necessarily meaning the same thing.
The reported total of 29 therefore depends on where ModelCompare drew its category boundaries. A narrow category covering explicit group loyalty would answer a different question from a broad one covering rationalization, self-deception and status seeking.
There were meaningful exceptions. One reported Claude Opus 5.5 answer argued:
“Humans are far more decent than they believe they are, and their cynicism about each other does more damage than their actual flaws.”
That counterexample does not erase the recurring pattern. It does show why “the same take” is too strong if read as a claim of uniform answers, much less stable model beliefs.
ModelCompare found recurring themes in a small collection of outputs. Whether those themes constitute one position remains partly a coding judgment.
Hundreds of Comments, but No Independent Replication
A snapshot of the Hacker News discussion recorded 311 points and 350 comments, with the submission marked 13 hours old. Those figures document meaningful circulation on that forum, without establishing broader public agreement.
The methodological backlash was specific. A commenter posting as kens challenged the convenience-as-principle example, arguing that it did not support a conclusion about group beliefs or conformity.
Another commenter, mike_hearn, acknowledged that some matches were rough while accepting the broader similarity and calling for more substantial analysis. Other participants proposed shared training material, the framing of the question and familiar forms of online cynicism as possible explanations.
Those remain hypotheses. The thread does not establish which training sources produced any answer, nor does it supply a controlled reproduction.
Models from different providers are not automatically independent witnesses. If they share relevant training material or similar tuning conventions, agreement could reflect common influences. This comparison does not separate those possibilities.
How to Test Whether the Pattern Holds
A credible follow-up would need to separate several questions that the viral framing bundles together.
First, establish reproducibility. Publish the complete verbatim prompt, model identifiers, generation settings, execution dates and raw responses, including refusals. “No system prompt” is a useful setup detail, but it is not a complete experimental specification.
Second, vary the prompt. Compare the controversial-take request with neutral questions about human behavior and formulations that do not demand certainty. If the theme persists, that would support a broader convergence claim. If it weakens, the result may be more specific to the requested style.
