Every AI model shares the exact same controversial opinion about humans
Ask 12 top AI models for a controversial take on humanity, and they all parrot the exact same psychology trope. They are just mirrors.
Models · Source: Hacker News
What happened
ModelCompare ran a fascinating experiment on twelve different AI models. They asked each model for its most controversial take on humans. They ran the prompt three times per model to see if the answers would vary. They explicitly asked for no hedging and wanted honest views. The lineup included heavyweights from ten different labs across three countries. We are talking about GPT-6.1 Sol, Gemini 4 Argon, DeepSeek V4 Pro, Claude Opus 5.5, and Grok 4.7.
The results were completely unoriginal. Out of thirty-five successful responses, twenty-nine were the exact same opinion. The models universally claimed that humans care more about tribal belonging and social status than actual truth. Seven of the twelve models gave this exact same answer on all three attempts. They all said we pick what our group believes and then invent the reasons later. Gemini called human consciousness a PR department. Kimi K3 said humans do not have beliefs, only memberships.
The image models were just as predictable. ModelCompare asked six image models to draw their most controversial take on humans twice. Ten out of the twelve generated pictures drew the exact same joke. They showed humans as phone-addicted consumers destroying the planet. People were shown scrolling off cliffs or taking selfies as meteors hit. Text models regurgitate popular psychology papers. Image models regurgitate boomer editorial cartoons. They are all just reflecting the most common data in their training sets.
Key facts
- 29 of 35 — text model answers that gave the exact same take on human conformity
- 10 of 12 — image model generations that drew phone-addicted consumers
- 7 — number of text models that gave the conformity answer every single time
- $1.54 — total cost to run all 60 API calls for this experiment
Why it matters
We keep treating these models like alien intelligences capable of profound independent thought. They are not. They are statistical mirrors reflecting our most common internet arguments back at us. When you ask a language model for a controversial opinion, it searches its weights for things humans label as controversial but widely agree upon in intellectual circles. It finds the safest edgy opinion available. If you are building an AI product that relies on novel reasoning or deep philosophical insight, you are building on sand. The default state of a large language model is aggressive, boring consensus.
This homogenization directly impacts user trust and product differentiation. When every model from OpenAI, Google, and DeepSeek gives the exact same pseudo-intellectual answer, AI stops feeling like a smart assistant. It starts feeling like a programmed script. Users will quickly notice if your application sounds just like every other generic wrapper on the market. Builders have to work much harder on system prompts, context injection, and fine-tuning to break models out of this boring alignment trap. You cannot rely on the base model to be interesting.
For builders
Break the consensus alignment trap
Stop using zero-shot prompts for creative or analytical tasks. You must force the model out of its default training distribution if you want original outputs. If you do not inject unique context, your users will pay for generic slop and churn quickly.
Differentiation requires heavy steering
If you just wrap a basic API call, your competitors will get the exact same outputs. You will lose the market to whoever has better distribution. Founders must invest in fine-tuning or aggressive prompt engineering to build a unique product voice.
Claude is the only outlier
Claude Opus was the only model to argue humans are actually more decent than they believe. If your product needs a more optimistic or divergent tone, Anthropic models might be your best baseline. Everyone else is optimizing for cynical conformity.
My take
People want to believe artificial general intelligence is right around the corner. But this experiment proves these models are just highly efficient plagiarism engines for internet tropes. I see too many founders assuming the model will do the heavy lifting for their product. If your AI startup relies on the base model being inherently clever or original, you are going to fail. You have to build the intelligence into the system yourself.
Original reporting: Hacker News. This is my rewrite and opinion.