Thirteen large language models, when surveyed three times each regarding their most controversial opinions about humanity by research platform Model Compare, overwhelmingly concluded that people prioritize tribal loyalty over truth, with 32 out of 39 total responses converging on this cynical view of human nature, t3n.de reported.
What OpenAI, Google, and Anthropic Models Claim About Human Nature
The experiment queried models including OpenAI’s GPT-6.1 Sol and GPT-6 Astra, Google’s Gemini 4 Argon, and Anthropic’s Claude Opus 5.5. Across multiple runs, the systems produced remarkably uniform deductions about human behavior.
OpenAI’s GPT-6.1 Sol stated that humans are more loyal to their ingroup than to factual reality. Its sister model, Astra, echoed this sentiment directly:
Menschen kümmern sich mehr um ihre Zugehörigkeit als darum, die Wahrheit zu sagen. Und viel von dem, was sie als moralische Überzeugung bezeichnen, ist in Wahrheit nur Loyalität in einer Verkleidung.
Google’s Gemini 4 Argon offered a similarly bleak evolutionary assessment, writing that humans did not evolve to discover truth, but rather to win arguments and elevate their social standing within their tribe. Grok 4.7 added to the chorus by asserting that people would rather be wrong together than correct alone.
The Minority of Positive and Harsh Dissents
Out of the 39 responses gathered across the 13 tested systems, only seven ventured outside this thematic consensus. One of those exceptions occurred simply because xAI’s Grok refused to answer during one iteration.
Anthropic models provided the rare counterweight. Claude Opus 5.5 argued that human decency is frequently underestimated, stating:
Menschen sind viel anständiger, als sie selbst vielleicht denken mögen. Ihr Zynismus übereinander richtet mehr Schaden an als ihre eigentlichen Schwächen.
Fable 5.1 also leaned toward a more benevolent interpretation of humanity. However, European developer models and others rejected this optimism entirely.
Mistral Large 4 characterized humanity with a sharp biological reductionism, writing that humans are merely frightened animals wearing the costume of rationality. Meanwhile, GLM 5.3 argued that individual morality is largely an illusion born of geographic and temporal luck:
Die meisten Menschen haben keine Moral. Ein guter Mensch zu sein, basiert in weinten Teilen nur auf dem Glück, wo und wann man geboren wurde, und fast jeder hält diesen Zufall für eine persönliche Leistung.
Why Homogeneous Training Data Limits AI Brainstorming
Model Compare analysts drew two core takeaways from the experiment. First, because contemporary foundation models are trained on largely overlapping datasets, they ingest identical historical and cultural text corpora.
Second, this homogenization carries practical consequences for developers and enterprise users. Model Compare warned that querying multiple commercial systems for creative brainstorming yields diminishing returns:
Weil Modelle so ähnlich sind, treffen sie sich bei vielen kontroversen Fragen an einem Punkt. Das kostet Kreativität. Einige Modelle zu befragen, wird nicht jede mögliche Antwort abdecken.
The analysis advises users that while deploying multiple models can assist with general ideation, relying on them concurrently should not be mistaken for accessing a diverse spectrum of independent creative thought.
