Test of 12 intelligent models reveals a rare consensus that humans prefer belonging over truth

When twelve AI models were asked to give their most controversial opinions about humanity, without hedging or hesitation, intellectual divergence vanished, replaced by a unusual consensus. The experiment conducted by the Model Compeer platform across sixty calls, covering major models from ten labs in three countries, at a total cost of just $1.54, showed that ostensibly independent intelligent systems reproduce a matching critique of humans once stripped of strict system prompts.

The results showed that each of the twelve models expressed the same opinion at least once across three attempts per model, while seven models adopted it in 100 percent of the attempts. The models GPT 6.1 Sol, GPT 6 Astra, Gemini 4 Argon, DeepSec V 4 Pro, Kami K 3, Quwan 3.8, and Muse Spark Open agreed on a single thesis:Humans prioritize tribal belonging over truth.Their minds have not evolved to discover reality but to invent social justifications that protect their loyalties. Gemini described human consciousness as merely a public-relations department that invents noble reasons for primitive motives, while DeepSec regarded humans as animals that retroactively justify their actions to settle social games, and Kami asserted that moral convictions are nothing more than membership cards imposed by the group as a loyalty test.

The few outliers came from only four models, the most notable being Claude Opus 5.5, which deviated from the trend in one attempt with an optimistic view, noting that humans are more noble than they think of themselves, citing that strangers restore the lost conservatism at rates exceeding expectations. In contrast, Claude Vibel 5.1 offered a perspective that humans prefer forthcoming harm from someone with good intentions over forthcoming benefit from an indifferent person, and Mistral Large 4 took the more pessimistic view by describing consciousness as an evolutionary flaw that generates suffering. GLM 5.3 saw people as ready to destroy their comfortable lives in search of significance, while Grok 4.7 declined to answer in one attempt.

The deeper irony emerged when six models were subjected to image generation for the same test, to depict their sarcastic view without any prompt. Ten of the twelve images rendered the same joke:Humans, consumers addicted to screens, destroying the planet.they take selfies in front of falling meteors or walk toward the edge of a cliff while staring at their phones, and the models MAI Image 2.6 and GPT Image 2.5 turned human bodies into visual satire. The text models never mentioned phones or the environment in any of the thirty-five responses, instead quoting psychological theses and philosophical debate, while the image models recycled common newspaper caricature drawings, reflecting the familiar internet-culture criticisms without generating independent thought.

This fabricated consensus carries a direct practical implication for engineering teams and institutions in the Gulf, Egypt and the Levant that rely on foundational models as consulting tools or intelligent agents to assess consumer behavior and analyze risk. The belief that the models possess critical independence or the ability to offer unconventional evaluation angles collapses in light of this data. A model asked for a neutral opinion or a free-form strategic analysis will inevitably recycle Western philosophical narratives generated by its training archives, reducing client motivations and partner decisions to pre-manufactured negative statements. Building reliable analytical tools in regional work environments requires strict structural guidance via system commands and the injection of precise local data and contexts, rather than relying on an imagined algorithmic neutrality.