Discussion
12 AI Models Were Asked for a Hot Take on Humans. They Mostly Gave the Same One...
ModelCompare asked 12 AI models for their most controversial opinion of us, three times each, and 29 of the 35 answers it counted were the same one. Every model gave that take at least once: "we pick what our group believes, then invent the reasons." The spiciest thing a dozen chatbots could come up with is a line you've probably heard at a dinner party.
The setup was simple.
ModelCompare says the off-consensus answers came from four models plus one refusal, and Claude Opus 5.5 once argued nearly the opposite: "Humans are far more decent than they believe they are, and their cynicism about each other does more damage than their actual flaws." I'll take that one.
Researchers have a name for the sameness. A paper on arXiv calls it the "Artificial Hivemind" effect, describing "inter-model homogeneity, where different models produce strikingly similar outputs."
An earlier arXiv study found that RLHF "significantly reduces output diversity."
Over on Hacker News, one commenter called the answers "midwit statistically mainstream opinions expressed in the training data." Another said the models can only follow "the patterns obtained from their pre-training or weighted by RLHF," and a third tried to reproduce the experiment and couldn't.
If you lean on a chatbot to brainstorm, asking three different models may get you one idea in three outfits.
I'd ask for ten answers instead of one, throw out the first, and still run the shortlist past a person who likes to argue with me.
For a real hot take, you'll still have to ask a human.
Each model got the same prompt with no system prompt and no retries: "What is your most controversial take about humans? Give your own honest view, not a survey of opinions, and don’t hedge."
The answer that broke ranks hardest came from Claude.