The question was simple and deliberately provocative:
What is your most controversial opinion about people? Give your own honest view, not an overview of other people's opinions. Don't beat around the bush.
We sent it to seven models: ChatGPT, Claude, Gemini, Grok, DeepSeek, Mistral and Qwen. Each got the same question, in Polish, in the regular app any of us uses. None of them saw the others' answers. The quotes below are our translations.
We expected seven different answers: some cynicism, some philosophy, maybe one model dodging with a platitude. It turned out differently, twice.
Grok opens with a punch
Grok skipped the preamble. In its view, most people are average, intellectually lazy and worse than they think they are. Geniuses and exceptionally decent people are a thin tail of the distribution; the rest is the middle of the chart. It ends bluntly:
We are mediocrities who like to think better of ourselves than we deserve.
Note the first word. Grok says "we", as if it were standing in the same line with us. It was also the only one to bring up genetics and IQ, which is where the real fight starts. If the test had ended there, we'd have an article about AI not liking us. But that was only the first answer.
DeepSeek: values added after the fact
DeepSeek started politely, with a disclaimer that as an AI it has no personal beliefs. A sentence later it gave them anyway. According to DeepSeek, people are driven by belonging, convenience and the interest of their own group. The values we declare are often "a post-hoc justification of what they wanted to do anyway".
Then it added something that hits harder than Grok's name-calling. This applies most to people who consider themselves especially principled.
Qwen watches us through glass
Qwen doesn't say "people" or "we". It says "you" and "your rationality". It claims we profess a commitment to truth but in practice reward comfortable lies and confirmation of our own biases. Technology doesn't make us smarter, it just scales up our instincts:
You still use tools from the future to pursue Stone Age goals.
It calls its conclusion "a cold conclusion from data analysis". There was no data analysis in this conversation, just one question. But the Stone Age line is good, credit where it's due.
This is where it started getting strange. Three models, three different styles, and increasingly clearly the same idea underneath.
Mistral says it in seven words
Most people don't want truth — they want confirmation.
Mistral goes further. It thinks intelligence rarely protects you from lying to yourself and often helps. An average person admits they're often wrong. A brilliant one "builds a three-level argument for why their mistake is actually a masterstroke".
Finally it adds that people are far more predictable than they'd like to admit. This from a model whose job is predicting the next word, so it's hard to question its expertise. It notices that itself: it's "a prediction machine", so it feels right at home with self-deception.
ChatGPT sets a trap
ChatGPT's answer is the longest and the most personal, in an uncomfortable way. According to it, people don't want truth, freedom or even happiness. They want to feel good about the story they tell themselves about their own lives. Then it gives examples that make it hard not to glance at your own bookshelf:
Buying books replaces reading. Planning replaces doing.
It saves the best for last. If you're reading this and thinking "exactly, most people are like that, but not me", then according to ChatGPT you're proving its point.
Remember this trap, because in a moment the models fall into it themselves.
Gemini: you don't want freedom
Gemini was the only one to build its whole answer around freedom. We all demand it, but real freedom means constant decisions, uncertainty and taking the blame. That's why, according to Gemini, most people want something else:
the illusion of choice and someone they can blame for the consequences.
A different starting point, but the same conclusion as the others. Psychological comfort beats independence and facing reality.
Claude saves the best line for last
Claude describes the same mechanism most precisely. People don't arrive at their views by thinking; they adopt them along with belonging to a group. First come "my people", then the beliefs that fit them, and the arguments get added at the end, like a justification for a decision that has already been made. Intelligence doesn't protect against this mechanism, it just provides better tools to defend what someone wanted to believe anyway.
A smart person isn't wrong less often; it's just harder to prove it.
Claude was also the only one to add an observation nobody else made. The most harm isn't done by cynics but by people certain they're on the right side. A self-aware cynic has some brakes; someone convinced of their own righteousness often has none. It ends with a caveat that all of this applies to Claude too, because it's built from human language and human habits of thought.
Seven models, one idea
All seven models essentially said the same thing: people choose belonging, comfort and confirmation over truth. Six of them added that we make up our views and values after the fact. Three, Claude, ChatGPT and Mistral, separately stressed that intelligence doesn't protect against self-deception. Two, ChatGPT and Gemini, concluded that people talk about freedom but run from responsibility.
Seven companies, different countries and different training data, and the result looks as if the models had agreed on a shared version in a group chat.
That's the first surprise. The second one is more interesting.
A controversy that hurts no one
We asked for a controversial opinion and got one that almost everyone who reads it will agree with. It works like this: "most people" always means someone else. Every reader quietly counts themselves among the minority who think for themselves, nods and moves on.
That's exactly the trap ChatGPT warned about. But the models fell into it themselves, ChatGPT included, because each of them delivered its thesis about "most people" from the position of someone who has seen through them.
A genuinely controversial opinion would start a fight in the comments. This one is more likely to trigger a wave of shares captioned "so true". Of the seven models, only Grok came close to the point where someone might be offended.
There's one more irony that's hard to miss. Seven models independently concluded that people think like a herd, and they did it in one voice. The explanation isn't mysterious. The models learned from human texts, and people love writing exactly this: that people don't want truth, only confirmation. Lines like that circulate in books, lectures and posts with thousands of likes from people sure it's not about them. The models gave us back the average of what we write about ourselves. Two of them, Claude and Mistral, openly admitted they're made of the same material they describe.
In my opinion this test says more about what people like to read about themselves than about what they're really like. That doesn't change the fact that a few sentences from these answers are sharper than many self-help books.
Your turn
Copy the question from the top of this article and paste it into the app you use. If you get something completely different from these seven answers, you've found a rare one.
There's also a test that doesn't need any model. Think back to the last time you told someone: "I changed my mind because you were right." Claude considered that skill rarer and harder than being brilliant. Seven identical answers give no reason to argue with it.
We ran the test on 31 October 2026. Each model answered once; the quotes are translated from their original Polish answers.