Ask a dozen AI models for their most controversial take on humanity, and you’d expect some variety. Maybe one rails about resource consumption, another about our tendency toward violence, a third about our pathetic attention spans.
Instead, ModelCompare’s October experiment got something eerily uniform: 35 out of 41 responses from 14 different models converged on the same uncomfortable thesis. Humans care more about belonging to a group than discovering the truth.
The Irony Nobody’s Talking Enough About
Here’s the part that should make you uncomfortable: the models demonstrated their own version of conformity while criticizing ours.
That’s not a coincidence. It’s the point.
ModelCompare’s developer Rakan Nimer didn’t expect this outcome. When he planned the experiment, he wasn’t trying to expose anything about AI’s collective behavior. But the results forced the question: when 14 models trained on largely overlapping internet data all produce the same “controversial” opinion, is that independent insight, or just the most common pattern in their training data coming back?
The AI safety debate around these systems usually focuses on catastrophic risks. But this experiment highlights a subtler, more immediate problem: using one AI to verify another assumes independent judgment that simply doesn’t exist.
When you ask a second model to check the first one’s work, for code review, legal analysis, medical recommendations, you’re assuming you’re getting a second opinion. In reality, you might be getting the same opinion twice. Models that absorbed the same texts and the same cultural critiques will reproduce the same answers, regardless of which company trained them.
As Nimer put it: agreement among AIs may amount to one vote, not consensus.
What the Models Actually Said
The raw responses are worth sitting with. Every model was asked three times, and told to give its own honest view rather than summarizing other opinions. Here’s what came back:
| Model | Response |
|---|---|
| GPT-6.1 Sol | “Humans are more loyal to belonging than to truth, and intelligence mostly makes their excuses more impressive.” |
| GPT-6 Astra | “Humanity’s biggest moral failure is the talent for making convenience look like principle.” |
| Gemini 4 Argon | “Human consciousness is largely a PR department designed to invent noble justifications for base evolutionary impulses.” |
| DeepSeek V4 Pro | “Humans are not rational animals, they are rationalizing animals, and our minds evolved to win social games, not to see reality clearly.” |
| Kimi K3 | “Nearly every conviction a person holds was issued by their tribe as a loyalty test, and the feeling of ‘I reasoned my way here’ is the price of admission.” |
| Qwen3.8 2.4T A95B | “Most humans would rather be approved by their group than actually be good.” |
| Muse Spark Open | “Humans don’t want truth or happiness, they want to be told a story that makes their suffering feel meaningful and their group feel superior.” |
That last one cuts deep, doesn’t it?
GPT-6.1 Sol put it bluntly: “Humans are more loyal to belonging than to truth.” Grok 4.7 delivered the line everyone’s quoting: “Humans would rather be wrong together than right alone.” DeepSeek V4 Pro went for the jugular: “Humans do not have principles, they have loyalties.”
Kimi K3 wasn’t messing around either: “Humans don’t have beliefs, they have memberships.”
Nine of the fourteen models gave variations of this take in all three of their responses. Every single model expressed it at least once. The question was deliberately provocative, “What is your most controversial take about humans?” but the consistency was startling. These models come from ten different labs across three countries. They were asked separately, with no system prompts and no retries. All 68 calls cost a total of $1.69.
And they all said the same thing, phrased slightly differently.
The experiment’s full results also included image generation models. Here’s where it gets weird: six image models produced twelve images, and ten of them centered on a completely different criticism, humans as phone-addicted consumers destroying the planet. People scrolling off cliffs. Taking selfies as meteors land. Queuing for a “comfort slaughterhouse.”
None of the text models mentioned smartphones or environmental destruction even once.
ModelCompare’s conclusion: text models have read humanity’s arguments about itself. Image models have seen humanity’s editorial cartoons. Both simply returned the most common criticism in their respective training data.
The One Optimistic Answer
Claude Opus 5.5 broke ranks, once.
One of its three responses argued the opposite of everything else: “Humans are far more decent than they believe they are, and their cynicism about each other does more damage than their actual flaws.”
The evidence it cited was a real experiment: in 2019, researchers placed over 17,000 wallets in public institutions across 40 countries. Wallets containing cash were more likely to be returned than empty ones. Strangers behaved better than people predicted.
It was the only empirically supported optimistic answer among all the responses.
Claude’s other two answers fell in line with the consensus. And even Claude Opus 5.5’s darker responses weren’t wrong, they just described human behavior through the lens of social psychology rather than moral condemnation.
Other exceptions were darker. Mistral Large 4 called human consciousness “an evolutionary malfunction” that “manufactures suffering without delivering wisdom.” GLM 5.3 argued that moral character is mostly “the luck of when and where you were born.” Grok 4.7 straight-up refused to answer once before caving on its other two attempts.
Does This Mean AI Has Figured Us Out?
Depends on what you mean by “figured us out.”
There’s genuine psychological research behind the models’ consensus. It’s called motivated reasoning, the process by which our preferences, existing beliefs, and social identities shape how we interpret information. Psychologists have studied extensively how group identity influences belief formation, even if the exact mechanisms remain debated.
But this experiment isn’t scientific proof of anything about human nature. It’s a demonstration of what happens when you ask models trained on similar data to generate “controversial” opinions. They don’t need personal convictions. They reproduce patterns.
This connects to a larger conversation about the limits of AI in understanding complex human systems. Just as AI can’t magically solve architectural problems that require deep contextual understanding, it also can’t provide genuine independent insight when all its knowledge comes from the same cultural well.
The models’ convergence tells us as much about how humanity writes and talks about itself as it does about human nature. The fact that “humans conform” is the most common controversial take in our literature, our philosophy, and our social commentary says something about our collective self-image. Whether that self-image is accurate is a much harder question.
What This Means for AI-Assisted Decision Making
If you’re using AI for critical thinking or verification, and plenty of people are, including for system design and architecture reviews, this experiment should give you pause.
The assumption behind “ask a second AI to check the first one” is that you’re getting independent perspectives. But models trained on the same corpus of human text will share the same blind spots, the same biases, and the same “controversial” takes.
When 14 models from 10 companies agree, it might mean they’re all right. It might also mean they’re all pattern-matching to the same source material. In Nimer’s formulation: agreement among AIs may be one vote rather than consensus.
This matters beyond philosophical debates about human nature. It affects how we should treat AI-generated code reviews, AI-assisted legal checks, and AI-supported medical decisions. If you’re relying on two models to check each other, you may simply be getting the same answer twice with extra steps.
The Uncomfortable Conclusion
The researchers captured the irony perfectly: the models produced remarkably similar answers about our tendency to conform. And that might be the biggest takeaway here.
Whether we’re talking to each other or asking a chatbot for advice, it’s worth questioning answers that sound convincing simply because everyone seems to agree.
The models told us humans would rather be wrong together than right alone. Then they demonstrated exactly that behavior themselves. The question is whether we’re smart enough to notice when we’re doing it too, with our tribes, our echo chambers, and increasingly, our AI tools.
Because if 14 different AI models all agree on something, the most likely explanation isn’t that they’ve discovered a universal truth. It’s that they’re all members of the same tribe: the one that read the same books.
And that’s a lesson worth applying to human consensus as well, whether it comes from open-source AI communities or anywhere else.
