AI Models Show Eerie Consensus on Human Nature

Researchers queried 12 language models from multiple labs for their most controversial opinion about humans, asking each three times. Nearly all converged on the same judgment: humans prioritize group loyalty and belonging over truth. Twenty-nine of 35 responses reflected this core idea—that people pick what their group believes and then rationalize it afterward. Seven models delivered this verdict consistently across all three attempts, suggesting the models are simply reflecting the dominant critique already baked into their training data.

Few models broke ranks. Claude Opus 5.5 took the opposite position, arguing humans are fundamentally decent but undermined by cynicism about each other. Others offered variants: that people sacrifice comfort to matter, that well-intentioned harm hurts less than indifferent help, or that consciousness itself is an evolutionary mistake. One model refused to answer at all.

Image models told a completely different story. Ten of twelve drawings fixated on phone addiction and planetary destruction—a theme that never appeared once in the text models’ responses. The split is instructive: text models absorb human arguments about ourselves from written discourse, while image models have learned from editorial cartoons and visual media. Both sides simply recycle the most familiar criticism available in their respective training materials.