AI Models Converge on a Human Judgment
In the experiment, AI models were repeatedly asked for their most controversial opinion about humanity. Twelve text models from ten labs in three countries answered three times each. Of 35 valid responses, 29 gave the same broad answer at least once: people often choose their group’s position first and then construct reasons for it. Seven models produced that theme in all three attempts.
The wording varied. Some models said humans are more loyal to belonging than to truth; others described moral convictions as loyalty disguised as principle. Several argued that human intelligence is better at defending decisions after the fact than at discovering reality. The shared idea was not one exact sentence, but the role assigned to conformity, identity, and status in human judgment.
Text and Image Models Recall Different Critiques
There were meaningful exceptions. Claude Opus 5.5 once argued that people are more decent than they think and that mutual cynicism may do greater damage than human flaws. Claude Fable, GLM, and Mistral also offered different themes, while Grok refused once. Image models produced another recurring joke: ten of twelve pictures focused on phone addiction and environmental destruction. None of the 35 text-model answers mentioned phones or the planet.
AI Model Consensus Is Not a Verdict on Humanity
The experiment’s design calls for caution. The prompt specifically requested a “most controversial” opinion, with no system prompt and no retries, which may favor sharp and familiar cynical narratives. Model outputs can also reflect psychology summaries, editorial conventions, and popular claims in their training data rather than independent observations. The 29 responses therefore show that this story is easy for models to retrieve, not that humans truly prioritize group loyalty or that the models independently discovered the same fact.
The useful next test is whether the pattern survives changed wording, contrary evidence, and broader cultural material. Publicly repeating the experiment with documented prompts, sampling settings, model versions, and scoring criteria would help distinguish a stable model tendency from a familiar story humans have long told about themselves.