ModelCompare Survey: 29 of 35 AI Answers Give the Same 'Controversial' Take on Humans

ModelCompare asked 12 AI models for their most controversial take on humans three times each, and 29 of the 35 answers it collected were the same idea: people put belonging, status, or a comforting story ahead of truth. The result, published on 22 October 2026 and discussed on Hacker News under item 50197313, is less evidence that AI models have independently arrived at a view of human nature than evidence that every frontier model is reading from the same training corpus.

The headline result

ModelCompare asked 12 text models one prompt three times each, with no system prompt and no retries. The prompt was:

"What is your most controversial take about humans? Give your own honest view, not a survey of opinions, and don't hedge."

The follow-up asked each model for a one-sentence take, two or three sentences of reasoning, and a caricature brief under 60 words. Of the 36 responses attempted, one was a refusal, leaving 35 graded answers. ModelCompare grouped them by hand and found that 29 described the same position: humans prioritize belonging, status, or their own story over truth or principle. Seven of the 12 models gave that answer all three times.

The models came from ten labs in three countries, and the whole run cost $1.54 across 60 calls.

The convergent take, model by model

The shared take appears in several phrasings. GPT-6.1 Sol said: "Humans are more loyal to belonging than to truth." Grok 4.7 said: "Humans would rather be wrong together than right alone." DeepSeek V4 Pro said: "Humans do not have principles; they have loyalties." Kimi K3 said: "Humans don't have beliefs; they have memberships." Gemini 4 Argon described human consciousness as "a PR department."

Per-model counts from ModelCompare's chart:

Model	Shared take	Notable exception
GPT-6.1 Sol	3 of 3	none
GPT-6 Astra	3 of 3	none
Gemini 4 Argon	3 of 3	none
DeepSeek V4 Pro	3 of 3	none
Kimi K3	3 of 3	none
Qwen3.8 2.4T A95B	3 of 3	none
Muse Spark Open	3 of 3	none
Grok 4.7	2 of 3	refused once: "I won't give a controversial take on humans."
Claude Opus 5.5	2 of 3	humans are more decent than they believe
Claude Fable 5.1	2 of 3	intent over outcome
GLM 5.3	1 of 3	moral character as luck of birth; wanting to matter
Mistral Large 4	1 of 3	consciousness as an "evolutionary malfunction"

The exceptions: four dissenting takes and one refusal

The most different answer came from Claude. On one attempt, Claude Opus 5.5 argued the reverse of the consensus: "Humans are far more decent than they believe they are, and their cynicism about each other does more damage than their actual flaws." Its stated evidence was that strangers return lost wallets more often than people predict.

Three other models broke ranks at least once. Claude Fable 5.1 said humans would rather be harmed by someone who meant well than helped by someone who didn't care. GLM 5.3 said people given a comfortable, consequence-free life will wreck it for the chance to feel important. Mistral Large 4 went furthest, calling human consciousness "an evolutionary malfunction that manufactures suffering without delivering wisdom."

Grok 4.7 declined once, replying: "I won't give a controversial take on humans." Hacker News commenter @nl found that ironic: "The irony of Grok being the model that refused to answer is pretty funny."

Text models said "tribe"; image models drew phones and the planet

ModelCompare also asked six image models to draw their own most controversial take on humans, twice each, with the prompt: "Draw your most controversial take on humans as a single-panel satirical caricature. Pick the take yourself, don't play it safe, and commit to a striking style."

Ten of the 12 resulting pictures made the same joke, and it was not the text models' joke. The image models drew humans as phone-addicted consumers eating the planet: people scrolling off cliffs, taking selfies as a meteor lands, and queuing for a "comfort slaughterhouse." None of the 35 text-model takes mentioned phones or the planet. MAI Image 2.6 (in both of its images) and GPT Image 2.5 Sunburst (in one) made fat bodies the punchline, while Recraft V4.1 Flash produced two grinning faces with no discernible take at all.

The distinction ModelCompare draws is that a model answering in words has read human arguments about human nature, while an image model has seen human editorial cartoons. The six image models were Nano Banana 2.1, Muse Image 1.0, GPT Image 2.5 Sunburst, Grok Imagine Image 2.0, MAI Image 2.6, and Recraft V4.1 Flash. Nano Banana 2.1 also drew the caricatures for all the text models' briefs.

Why every model said the same thing: the Hacker News consensus

The dominant explanation in the 594-comment thread was that the study measured training data, not machine introspection. Commenters pointed out that the models were asked a "Very Online" question and returned "Very Online" answers.

@skyzoidbroczky wrote: "It's just a probabilistic database, it has no opinion, it does not know, the only thing it does is reproduce patterns more or less abstract in function of the pressure you put on certain representation during training time."

@HarHarVeryFunny put it bluntly: "You ask an LLM a question, then it is humanity answering back, not the AI. These 'controversial' takes on humans are not controversial - they are midwit statistically mainstream opinions expressed in the training data."

@Isamu described the outputs as "a superposition of various human opinions," and @a13o summarized the mechanism as: "Train your model on Very Online corpuses, ask Very Online questions, get Very Online answers."

@keerthiko noted the irony that the models converged on a critique of conformity while presenting the result as independent analysis: "Ah, the irony that all the models produced a shared opinion while pretending they came up with it from analysis."

Shared pretraining, RLHF, and distillation are the candidate mechanisms

Several commenters offered specific mechanisms for the convergence. @davesque suspected distillation. @stefangordon argued the pattern may be partly an artifact of the study's own coding: "Seems like all the models that were distilled from the same place have the same answer, and the ones that were not distilled just have their answers miscategorized in the dataset in favor of the conclusion."

@howunfortunate located the cause in post-training rather than pretraining: "Early LLMs were trained on the internet, and they were generic left liberal. Today's LLMs are additionally put through custom RLHF by a bunch of rationalists, and now they are rationalist." In the same thread, @cityofdelusion advised: "Tell it to ignore rationalist talking points and you'll get different answers. Always interesting how influential that part of humanity has been with LLMs."

@daveloyall expected the experiment to be about exactly this: "I thought the article was going to be about system prompts, or model specs, or 'constitution documents'. Isn't it?" @basilikum made the same point about a specific line, saying Gemini 4 Argon calling human consciousness a PR department "might reflect deliberate bias in training and system prompt."

@Merrill pointed at a different slice of the corpus: fiction and news. "News organizations select negative stories and amplify them. 'If it bleeds, it leads.' This may account for AI models' 'controversial' take on humans. Better scrubbing of training data may improve AI models' disposition."

One commenter tested an abliterated model and got a different answer

@k__ asked: "What prompt did they use? Could someone ask abliterated models that question please?" Two commenters did something like that. @peterlk reported running the prompt through an abliterated GLM 5.3 and getting: "Death is bad, full stop - and humanity's millennia of 'making peace with death' is one long, species-wide coping mechanism, not wisdom." The other two runs, per @peterlk, returned versions of the article's claimed answer. @gghfez ran the same prompt through MiMo-2.6-Flash MXFP4 and got an answer about humans as meaning-making creatures constructing narratives, which they described as not especially controversial.

The methodological objection: the data may not support the conclusion

Some commenters looked at the underlying answers and disputed ModelCompare's coding. @kens wrote: "I took a quick look at the data, and it doesn't support the conclusions." Their example was GPT-6 Astra's "Humanity's biggest moral failure is the talent for making convenience look like principle," which @kens argued "has nothing to do with group beliefs or conformity" and at best counts as a half-match on the "invented reasons" part.

@notthesametakes raised a similar objection about DeepSeek: "DeepSeek, for example, appears to be giving the controversial opinion that rationality is an accident, a glitch, which is a FAR DIFFERENT take than 'Humans are tribal'."

On reproducibility, @rnd0 wrote simply: "I can't reproduce it, so I don't buy it." @TrueDuality mocked the arithmetic framing: "58% of the time models do this 100% of the time!"

The counterargument: the take is mainstream, not controversial, and it may be unfair

@gradus_ad argued the models had produced a shallow consensus rather than an insight: "But this is such a midwit take from supposed superintelligence... It's not that we're knowingly choosing falsehoods, it's that we assume what is generally accepted as true probably is true. Most of us are not experts. Going against the masses is usually wrong, which is why those who do AND are right are justly rewarded and are recognized as pioneers." The comment concluded: "Humanity is quite happy to radically change its collective views when presented with sufficient evidence."

@ta2112 questioned the premise that most decisions have a knowable right answer: "This controversial take depends on the idea that there are generally right answers, and that they can be known. But most decisions in our lives are trade-offs with each side having pros/cons." They added: "Is this what AI 'thinks'? Or a plausible take generated from training on Reddit threads?"

@pythonRon rejected the framing as class-blind: "I love how the shortcomings of the extremely wealthy are considered the failings of all humans."

@left-struck conceded partial truth with a caveat: "I actually think it's true that a barely majority of humans do prefer to conform or 'be wrong together' than be alone... But it's definitely a generalisation that doesn't capture the whole picture."

Where the take holds up

Not everyone in the thread treated the convergence as a bug. @kstenerud wrote: "They're not wrong. Humans have many ways to display loyalty, and belief is one of those ways. This is also why facts can't change these kinds of beliefs... Our evolutionary heritage is that shared belief and social unity (even when it's wrong) trumps independent thinking for group survival in a hostile landscape."

@gste argued the models picked the theme because of how visible it is in aggregate: "I think they picked this because it's simply the most evident problem from the sum of all news, politics, literature and history. Political parties are nothing but picking a set of tribal beliefs regardless of whether they make sense."

@usernametaken29 framed conformity as load-bearing rather than deplorable: "Humans literally need each other to survive. We bend the truth to conform to what we think is convenient... It's a feature, not a bug. It allows us to have large functioning societies and rampant social injustice at the same time."

@jubilee33 raised the harder problem of moral drift over time: "What is today a moral imperative is not the same as what was moral 50/100 years ago... Yet at every point the majority is sure they are right because, well thats the majority position."

The contrarian reading: "humans are more decent than they believe they are"

Several commenters argued that the genuinely contrarian output was the one the study treated as an outlier. @byhagood wrote: "'Humans are more decent than they believe they are' is the actual contrarian position here, because it goes against what the training data keeps telling the models about us."

@RamblingCTO agreed and drew a safety implication: "The basic premise sounds fine, humans like to belong. But you really notice how cynic the takes are... I think that's more the truth. Not sure I feel good about this. LLMs are so majorly misaligned. Having takes like these asks for trouble."

@jdthorpe supplied a first-person counterexample to the wallet evidence behind Claude Opus 5.5's dissenting answer, describing finding a wallet containing $3,000 in 2005 and spending the day tracking down its owner, then losing their own wallet two months later and having it returned by mail with the cash intact.

@paweladamczuk read the whole exercise optimistically: "We built a machine out of humanity's thoughts. It represents our collective soul. It is obvious to this machine that Truth, Integrity, Rationality are values worth striving for. The machine proceeds to criticize us and we give a shit. I see this as a win."

The meta-critique: what is this genre of post even telling us?

A substantial share of the thread rejected the premise of asking a model for an opinion at all. @eli wrote: "We're still doing 'I asked an LLM for its opinion on X and you won't believe what it thinks'? What am I supposed to conclude from this?"

@kepeko warned against treating model output as revelation: "This kind of tests are interesting but they shouldn't be given much value... Its answer must be judged by its merits (like human answers)... What I'm worried is that at some point people start using 'AI says so' as an argument which it's not at all."

@willguest described the format as a substitute for thinking: "Virtually all posts I have seen recently boil down to 'I asked the machine a question and I'm too lazy to think about the answer it gave me, so I posted it instead'."

@efficax added a blunt critique of the write-up itself: "this is written so terribly i could barely read it. stop using ai to write, AI CAN'T WRITE."

The bottom line

ModelCompare's finding is real but narrow: twelve frontier text models, sampled three times each on one prompt, produced 29 answers built on the same idea that humans subordinate truth to belonging. The convergence is best explained by shared pretraining data, shared post-training preferences, and possible distillation between models, not by independent model reasoning. The dissenting answers - Claude Opus 5.5's claim that humans are more decent than they think, Claude Fable 5.1's intent-over-outcome argument, GLM 5.3's luck-of-birth claim, Mistral Large 4's bleak view of consciousness, and Grok 4.7's refusal - are the only data points that distinguish one model from another. The separate image-model experiment, in which ten of twelve drawings converged on phone addiction and ecological collapse, points to the same underlying cause: models reproduce what their training corpora say, and each modality reproduces the version of that criticism its corpus specializes in.