Recently, ModelCompare released the results of an AI experiment in which researchers posed questions to 12 AI models, including ChatGPT, Claude, Gemini, Grok, and DeepSeek, asking for their views on humanity's most controversial truths. The responses from all models exhibited a high degree of consistency. During the experiment, researchers asked each model the same question three times, explicitly requesting the models to output their own opinions rather than summarizing others' views. A total of 35 text responses were collected, of which 29 expressed the same core idea: compared to seeking the truth, humans care more about group belonging.
AI Models Agree That Humans Value Group Belonging
Specifically, GPT-6.1 Sol suggested that humans prioritize group belonging over truth, noting that intelligence is often used to make one's excuses more persuasive; Grok 4.7 stated, "Humans would rather be wrong together than right alone"; Gemini 4 Argon believed that humans use intelligence to justify instinctual decisions rather than derive objective conclusions; DeepSeek V4 Pro pointed out that humans are better at rationalizing their beliefs than examining those beliefs themselves. Statistics show that seven models provided variants of these criticisms in all three responses, and all models tested mentioned related viewpoints at least once, with only differences in wording.
Only a few responses presented differing viewpoints. Claude Opus 5.5 suggested that humans are actually kinder and more decent than they perceive themselves to be, noting that while people sometimes give negative evaluations of one another, strangers often exhibit generous behavior. However, this model also referenced the general criticism of group conformity in its other two responses. Additionally, Mistral Large 4 described human consciousness as an evolutionary flaw, while another model mentioned that humans might sacrifice happiness to gain a sense of importance.
AI Image Generation Models Reveal Environmental Impact of Tech Consumption
The experiment also invited six AI image generation models to participate in a similar test. Among the 12 generated artworks, 10 centered around another core theme: that humans addicted to tech consumption are destroying the Earth. The content predominantly depicted humans engrossed in their smartphones and consuming resources, oblivious to the deterioration of their surroundings. This theme did not appear in the core criticisms from the text models. Researchers noted that the results of the experiment do not constitute scientific conclusions, as the viewpoints provided by the models stem from existing human discourse in their training data. Factors such as incentivized questioning, limited sample sizes, and manual categorization of answers all indicate that the results cannot directly prove that humans indeed value belonging over truth.
It is noteworthy that while several text models criticized human conformity, they themselves provided highly convergent responses.

