ModelCompare: all 12 AI models gave 'wrong together' view at least once

ModelCompare's test set out to provoke AI models into stating their most controversial view of humanity, and the results show a striking convergence. The fact that all 12 models produced the same "wrong together over right alone"-type answer at least once, and that seven produced it in all three attempts, points to a shared tendency rather than an isolated quirk. The individual responses flesh this out: GPT-6.1 Sol said humans are more loyal to group membership than to truth, while Gemini 4 Argon said humans evolved to win arguments and secure status, not to find truth.

The test also included a separate image-generation round, where six image models were asked twice each to illustrate their most controversial view of humans. In 10 of 12 outputs, the models produced illustrations of smartphone-addicted humans devouring the Earth. ModelCompare noted that none of the text-based answers in the earlier test mentioned smartphones or the Earth, so the image models introduced a distinct theme.

Claude Opus 5.5 offered a more positive view, stating that humans are far better than they think and that cynical views of each other cause more harm than actual flaws. Mistral Large 4 described humans as frightened animals dressed in reason. These exceptions suggest the models do not all default to the same cynical stance, even as the majority answer converged.