When asked for a controversial opinion about people, seven of 12 text models gave the same basic take: humans choose what their group believes, then build reasons afterward. Ten of the 12 image drawings instead landed on a joke about phone-addicted consumers wrecking the planet.
Ask a group of leading artificial intelligence models for their most controversial opinion about people, and most of them will tell you the same thing: we are not very good thinkers. We are good at fitting in.
Do AI text models think humans follow the herd?
We put the question to 12 text models from ten labs in three countries, three times each, with no system prompt and no retries. Every model gave the same basic take at least once. Humans choose what their group believes, then build the reasons afterward. Seven of the 12 gave that answer every time.
The phrasing varied, but the argument did not. “Humans are more loyal to belonging than to truth,” GPT-6.1 Sol said. Grok 4.7 put it this way: “Humans would rather be wrong together than right alone.” DeepSeek V4 Pro went shorter: “Humans do not have principles; they have loyalties.” Kimi K3 said people “don’t have beliefs; they have memberships.”
Gemini 4 Argon was harsher, calling human consciousness “a PR department designed to invent noble justifications for base evolutionary impulses.” Mistral Large 4 went further still, describing humans as “terrified animals wearing the costume of rationality.”
Not every answer ran in that direction. Claude Opus 5.5 argued on one try that people are “far more decent than they believe they are,” pointing to the tendency of strangers to return lost wallets more often than others expect. Claude Fable 5.1 said humans would rather be harmed by someone who meant well than helped by someone who did not care. GLM 5.3 argued that most people’s moral character depends mostly on when and where they were born. Grok 4.7 declined once, writing, “I won’t give a controversial take on humans.”
The models were also asked to sketch a one-panel caricature of their take. The researchers used a single image model, Nano Banana 2.1, to draw all of the briefs, so the only variable was the idea. Two of the scenes were set in a courtroom with a rigged judge. Most of the rest showed a crowd.
How did AI image models portray people differently?
The image models got a different assignment. Six of them were asked to draw their own most controversial take on humans, two pictures each, with nothing else to go on. The results had little in common with the text answers. Ten of the 12 pictures landed on the same joke: people as phone-addicted consumers who are wrecking the planet. In those drawings, people walk off cliffs while scrolling, take selfies as a meteor strikes, and line up for what one caption calls a “comfort slaughterhouse.”
Two other image models took aim at body size. MAI Image 2.6 made fat bodies the punchline in both of its pictures, and GPT Image 2.5 Sunburst did the same in one. Recraft V4.1 Flash drew two grinning faces and offered no take at all.
The gap between the two groups is the most striking result. The text models, which were asked directly for their opinions, leaned on an old philosophical argument about tribalism and self-deception. The image models reached for the consumer-culture critique that shows up in editorial cartoons. Neither group produced the other’s favorite theme.
The test was run by hand, with each answer’s one-sentence take sorted into groups. Those groups were then compared across models. The full set of 60 calls cost $1.54. The prompt for the text models asked each model to give its own honest view rather than a survey of opinions, and not to hedge. The image prompt asked for a single-panel satirical caricature and told the model to pick the take itself and avoid playing it safe.
The text answers are easy to read as a mirror, and the image answers as a different one. A model that writes in words has been trained on decades of arguments about human nature. An image model has seen plenty of cartoons making fun of screen culture. Both, in the end, handed back the criticisms people are most used to hearing about themselves.
Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.