Most of us would like to believe we're independent thinkers who make decisions based on facts, logic and personal values. But according to some of the world's biggest AI models, we're not nearly as independent as we think.
In a new experiment published by ModelCompare on October 22, researchers asked 12 AI models, including ChatGPT, Claude, Gemini, Grok and DeepSeek, to share their most controversial opinions about humans.
The results were wildly consistent. Of the 35 responses collected, 29 expressed essentially the same idea: humans care more about belonging to a group than discovering the truth.
Perhaps most ironically, the AI models seemed to demonstrate their own version of conformity while criticizing ours.
What the AI models said about humanity
The researchers asked each model the same question three times, instructing it to give its own view rather than summarize other opinions.
Some of the responses were particularly blunt. For instance:
GPT-6.1 Sol argued that humans prioritize belonging over truth, suggesting that intelligence often serves to make our excuses more convincing.
Grok 4.7 delivered one of the most memorable responses: "Humans would rather be wrong together than right alone."
Gemini 4 Argon went further, suggesting that people use their intelligence to defend decisions made by instinct rather than arrive at objective conclusions.
DeepSeek V4 Pro was similarly critical, arguing that people are better at rationalizing their beliefs than examining them.
Interestingly, seven models gave variations of the shared criticism in all three responses. Every model expressed it at least once. Although the wording differed, the message was largely the same. We like to think we're making independent decisions, but our desire for acceptance may influence our beliefs more than we realize.
Claude had a surprisingly different answer
Not every response was pessimistic. Claude Opus 5.5 offered one of the experiment's most unexpected takes, suggesting that people are actually kinder and more decent than they give themselves credit for.
Rather than criticizing humanity for selfishness or conformity, Claude pointed to the gap between how poorly people sometimes view one another and how generously strangers can behave.
That was a notable departure from the other responses, although Claude also echoed the broader criticism in its remaining two answers.
Other exceptions were considerably darker. Mistral Large 4 described human consciousness as an evolutionary malfunction, while another model suggested people might sacrifice happiness simply to feel important.
These exceptions helped illustrate that the models weren't incapable of producing different opinions. They simply gravitated toward a remarkably similar criticism.
AI image generators had a different complaint
The experiment didn't stop with chatbots. ModelCompare also asked six AI image generators to illustrate their most controversial opinions about humanity.
Ten of the 12 generated images centered on a different theme: humans as technology-obsessed consumers damaging the planet. The images depicted people scrolling on their phones while ignoring danger, consuming resources and remaining distracted as the world around them deteriorated.
Surprisingly, none of the 35 text-model responses mentioned smartphones or environmental destruction as their central criticism.
That difference may say something about the information each type of AI learns from. Language models are trained on enormous quantities of written material, including arguments about human psychology and behavior, while image generators learn visual patterns that can include familiar editorial cartoons and social commentary.
Does this mean AI has figured us out?
Well, I wouldn't say that, exactly. There is a genuine psychological concept behind the responses called motivated reasoning, which describes how our preferences, existing beliefs and social identities can influence the way we interpret information.
Research has also explored how strongly group identity shapes what people believe, although psychologists continue to debate exactly how these effects work.
But this experiment is not scientific proof that humans value belonging more than truth.
The models were given a deliberately provocative prompt, and the researchers grouped the answers manually into broad categories. The experiment involved a relatively small number of responses and didn't test how the results might change under different instructions.
AI models don't need personal convictions to generate forceful opinions. They can reproduce arguments and patterns learned during training. So, in other words, these answers may tell us as much about the way humanity writes and talks about itself as they do about human nature.
Final thoughts
What makes this experiment fascinating, at least in my opinion, is that so many different models reached for a familiar criticism when asked to say something controversial.
The researchers captured the irony perfectly: the models produced remarkably similar answers about our tendency to conform. And that might just be the biggest takeaway here. Whether we're talking to other people or asking a chatbot for advice, it's worth questioning answers that sound convincing simply because everyone seems to agree.