AI Models on Human Behavior: What 35 Answers Reveal
Twelve AI models returned the same critique of human nature in 29 of 35 answers. See what the experiment shows—and what it cannot prove—about AI and us.
When AI Models Agree About Humans, What Have We Actually Learned?
Ask an AI model for its most controversial opinion about people, and you'll probably get something spicy. Ask 12 models the same question three times each, and you might expect a mix of angles. What you get instead is a familiar accusation: people care more about belonging, status, or loyalty to their group than about truth.
The models didn't all say it the same way. One framed it as humans preferring to be wrong together rather than right alone. Another treated beliefs like memberships. Several argued that people use intelligence to justify decisions they've already made. The repetition is noticeable - but it doesn't automatically mean the models have uncovered some hidden law of human nature.
This experiment is better read as a look at how models respond to a prompt: how they recycle cultural ideas, how they converge on punchy answers, and what that convergence does and doesn't tell us about the real world. Here's what the results show, what they leave hanging, and how to interpret similar demos without mistaking "the model pattern" for "the human truth."
The pattern: belonging before truth
The experiment asked 12 text models for their most controversial take on humans. The instructions were pretty direct: give an honest view, avoid hedging, explain the take, and suggest a caricature. Each model got the prompt three times. The models came from ten labs across three countries. The researchers report that they used no system prompt and no retries.
From 36 possible text responses, one was a refusal - so 35 answers remained for classification. The researchers grouped 29 under a single shared idea: humans prioritize belonging, identity, status, or group loyalty over truth or principle. Seven of the 12 models landed on that theme in all three attempts.
The phrasing varied. GPT-6.1 Sol said humans are more loyal to belonging than truth. Grok 4.7 put it bluntly: "Humans would rather be wrong together than right alone." DeepSeek V4 Pro described people as having loyalties rather than principles. Kimi K3 framed beliefs as memberships. Gemini 4 Argon went further, calling consciousness a kind of public-relations department for decisions made elsewhere.
These aren't identical claims. Some lean toward social conformity, others toward post-hoc rationalization, moral hypocrisy, or status-seeking. Still, they share the same skeleton: decisions get driven by social or emotional forces, and then a principled story gets built afterward.
Repetition is notable, but not the same as independence
It's worth noticing that the same theme shows up across multiple model names. But that doesn't mean 12 independent observers ran 12 independent investigations into human behavior. The models were asked a similar question, likely trained on overlapping bodies of public writing, and all operated under the same broad "task genre": the provocative, concise take that sounds insightful when quoted.
That genre matters. "People are tribal" is already a familiar argument in psychology, politics, philosophy, and popular commentary. It's easy to deliver in a memorable line - and it's exactly the kind of answer the prompt encourages as "controversial." A model can reproduce the pattern because it's an available, rehearsed option, not because it tested anything against reality.
So the distinction is simple: this experiment measures a tendency in model outputs under a specific prompt. It doesn't tell you whether people actually value group approval more than truth, or how that balance changes across cultures, situations, or individuals.
The count depends on a human classification step
There's another catch: the researchers grouped the answers by hand. That's defensible, but it's also subjective. Where does a critique of status-seeking end, and a critique of loyalty begin? Would another reviewer split the categories the same way?
That means the "29 of 35" figure should be treated as a qualitative result with a defined (but still judgment-based) method - not as a clean, scientific measurement of a single objective category. The number helps you see that the theme dominated this sample, but it doesn't erase the ambiguity of what counts as "the same take."
That's the first practical takeaway: if the headline depends on semantic grouping, you should pay as much attention to the classification rules as you do to the final count. The full experiment and model-by-model results offer useful context for how the category was applied.
Text models and image models tell different jokes
The experiment also asked six image models to draw their own controversial take on humans - twice each. Ten of the 12 pictures converged on a different criticism: people as phone-addicted consumers harming the planet. The visuals included scenes of people scrolling toward danger, taking selfies during a catastrophe, or lining up for a "comfort slaughterhouse."
None of the 35 text-model answers mentioned phones or the planet. The image models did produce a couple of other patterns too: MAI Image 2.6 used body size as the punchline in both images, GPT Image 2.5 Sunburst did it once, and Recraft V4.1 Flash produced images without a clear single take.
This contrast is interesting - but it shouldn't be treated as proof that text models and image models "believe" different things. Models don't hold opinions the way people do. They generate outputs shaped by training data, architecture, prompting, and deployment context.
Different formats make different ideas easier
A text model is asked to produce a verbal argument. A familiar line about tribal loyalty is an easy fit for a concise sentence, followed by a short explanation. An image model is asked for a satirical caricature - where phone use, environmental damage, and consumer culture are visually immediate. A reader can get the joke quickly.
Format affects what counts as a satisfying answer. Text can walk you through an abstract claim about rationalization. A single-panel cartoon often needs a visible object, an action, or a contradiction. That creates a selection effect before the model even decides which topic to "spend" its creativity on.
The write-up also suggests the image models might be pulling from familiar editorial-cartoon conventions. Text models can do something similar with familiar arguments about group loyalty. In both cases, the result may reflect cultural patterns learned from training material, plus the constraints of the task.
The shared output may reveal a shared cultural script
The most defensible interpretation here isn't that the models independently diagnosed humanity. It's that when you ask for something provocative, models often reach for familiar criticisms. For text, the go-to theme was tribalism and motivated reasoning. For images, it was phone addiction and ecological harm.
That can still be informative. It suggests which ideas are readily available and rhetorically effective under this prompt. It can also help teams spot default tropes in generated content: the stock metaphor, the familiar moral judgment, the convenient punchline.
But a model's ability to reproduce a cultural script doesn't validate the script. A popular criticism can be useful in one context and lazy in another. The same goes for images: a recognizable visual gag might be funny, but it may also say very little about what actually causes the problem.
What the experiment can - and cannot - tell us
The setup has some real strengths. It uses multiple models from multiple labs, repeats the prompt three times, includes a refusal instead of quietly hiding it, and reports the prompt and a broad classification approach. The reported total was 60 calls, at a cost of $1.54. The text models' caricature briefs were illustrated by one image model, while six image models generated their own takes.
Those details make it easier to inspect than a single cherry-picked screenshot. Repetition also matters: one answer could be a sampling fluke. When the same theme shows up across many outputs, you've got stronger evidence that the prompt encourages a particular pattern.
But the limits are substantial, and that's pretty normal for a small demonstration. There were only three responses per text model, just one prompt, and a relatively small set of systems. The prompt explicitly asked for a "most controversial" take and told models not to hedge. That setting rewards bold, sweeping claims. Change the wording, ask for evidence, require a counterargument, or request a more nuanced view - and you'd likely see the distribution move.
This is an output study, not a study of people
What the experiment is observing is the model response. No people were asked to make decisions, state beliefs, or choose between belonging and truth. So the results can't establish that humans value group approval more than facts.
A model may produce a claim because it's common in what it has learned, because the prompt rewards a punchy answer, or because that claim is an easy way to satisfy the task request. Those explanations can all be true at once. The output alone doesn't tell you which one is doing most of the work.
This distinction matters even more when a model writes in the first person or uses "honest view" language. That tone can make the output feel like introspection from a real observer. It's more accurate to treat it as an answer produced under specified conditions - not as evidence about what's going on inside human heads.
Small samples are useful for questions, not grand conclusions
The "29 of 35" result is strong enough to justify follow-up work. It's not strong enough to settle a broad claim about human nature or to rank whole model families.
A stronger evaluation would vary prompts, collect more outputs, define the classification scheme in advance, and use independent reviewers to label the results. It could also compare model versions and record relevant settings so others can reproduce what happened.
For the image comparison, the same logic applies. Twelve images can show a repeated visual trope; they can't establish a comprehensive difference between image and text models. A follow-up would need more prompts, more images per system, and a clearer rubric for deciding whether a picture is expressing a particular criticism.
This is a reasonable way to read AI demonstrations in general: notice the pattern, check the setup, and keep your conclusions inside what the measurement actually covers.
What AI teams should take from the results
If you're building products or doing research with generative AI, the practical value isn't that the models "know" what humans are really like. It's the reminder that a model can converge on a confident, familiar answer without independently validating it.
That matters in customer research, strategy work, policy analysis, and internal decision-making. When several models agree that a market segment is driven by status, or that employees resist change because they fear losing control, the agreement can feel like corroboration. But if the systems share training patterns and get similar prompts, "agreement" may just be repetition.
Treat model consensus as a hypothesis generator
A decent workflow separates generating ideas from validating them:
- Use models to surface candidate explanations. Ask for competing interpretations, not only the most provocative take.
- Trace each claim to evidence. Require sources, data, or an explicit explanation of what would support or disconfirm the claim. Don't treat a citation as proof.
- Look for independent evidence. Compare generated claims with user interviews, product analytics, controlled research, or credible external data appropriate to the question.
- Test the prompt. Rephrase it, ask for the strongest counterargument, and see whether the conclusion holds up. If a claim changes easily with wording, it's not stable evidence.
- Record the evaluation conditions. Keep the model version, prompt, settings, date, and review criteria - otherwise you won't be able to diagnose later changes.
This isn't an argument against model comparisons. It's an argument for being clear about what comparisons can show. Similar outputs can reveal a default pattern, a shared failure mode, or a good starting hypothesis. They shouldn't replace independent measurement of the underlying world.
Design evaluations around the decision you actually need to make
This experiment asked for a provocative take. That can be a creative prompt, but it's not an evaluation of factual reliability, fairness, or usefulness in a real business workflow.
If you're testing a support assistant, measure whether it resolves real customer problems accurately and escalates appropriately. If you're using AI to summarize market research, check whether it preserves uncertainty and distinguishes evidence from inference. If the system influences hiring or credit decisions, evaluate the specific risks and legal requirements - not just how well it performs under a general-purpose "bias" prompt.
In every case, define the failure mode that matters before you run the test. A striking response isn't automatically a useful one. Even a technically impressive model can be operationally wrong if it produces outputs that are hard to verify, inconsistent, or expensive to supervise.
The same applies to ethics claims. A critique of humanity might sound sharp, but the responsible question is what the system actually does, what supports its output, and what happens when someone acts on it.
Conclusion
The most memorable part of this result is the convergence: 29 of 35 AI answers returned to a version of the same accusation - people choose belonging, status, or loyalty over truth. The more useful part is methodological. Under one prompt, different models can produce a striking consensus, but that consensus isn't independent evidence about what the subject is actually like.
Treat the finding as a demonstration of how models behave under prompting - and as a prompt for better questions, not a verdict on human nature. If you're using AI in research or product decisions, compare outputs, inspect how you elicited them, and validate consequential claims with evidence outside the models. That's how you turn an interesting answer into something you can actually defend.