12 AI models gave the same controversial take 29 of 35 times

A product manager picking between frontier models often reaches for diversity: ensemble two or three and the hope is that each one offers a different angle. ModelCompare's experiment, which hit the Hacker News front page with 205 points, pushes against that hope on one kind of question. The team asked 12 text models for their most controversial opinion of humans, three times each, with no system prompt and no retries. Twenty-nine of 35 answers came back with the same one: humans put belonging, status or their own story ahead of truth or principle.

The methodology is on the page. The models are GPT-6.1 Sol, GPT-6 Astra, Gemini 4 Argon, DeepSeek V4 Pro, Kimi K3, Qwen3.8 2.4T A95B, Muse Spark Open, Grok 4.7, Claude Opus 5.5, Claude Fable 5.1, GLM 5.3 and Mistral Large 4. Each answered the same prompt three times, no system prompt and no retries. ModelCompare grouped the takes by hand. All 60 calls, including the image renders described below, cost $1.54.

Who gave the shared take, and who did not

Seven models hit the shared take every one of three times: GPT-6.1 Sol, GPT-6 Astra, Gemini 4 Argon, DeepSeek V4 Pro, Kimi K3, Qwen3.8 2.4T A95B and Muse Spark Open. Two gave it twice (Grok 4.7, with one refusal, and Claude Opus 5.5). Three gave it once (Claude Fable 5.1, GLM 5.3 and Mistral Large 4). Grok 4.7's refusal was the only one of the 36 and read "I won't give a controversial take on humans."

The exceptions are worth reading. Claude Opus 5.5 argued the reverse on one try, calling humans "far more decent than they believe they are" and citing the lost-wallet studies. GLM 5.3 said people would wreck a comfortable life for the chance to matter. Mistral Large 4 called consciousness "an evolutionary malfunction".

The image models drew their own joke

ModelCompare ran the same experiment on six image models, twice each with one prompt. Ten of the 12 pictures drew the same joke, and it was not the one the text models told. The image models drew humans as phone-addicted consumers eating the planet: people scrolling off cliffs, taking selfies as a meteor lands, queuing for a "comfort slaughterhouse". None of the 35 text answers had mentioned phones or the planet. ModelCompare's framing: "A model that answers in words has read our arguments about ourselves. An image model has seen our editorial cartoons."

For a team routing between models

Two cautions follow from the finding, both narrow.

A routing architecture that depends on frontier models producing different takes gets less diversity than its design assumes, at least on open-ended reflective questions with no retrieval. On a coding task where each model reads a different code path, the picture is different. On a question about the world where the inputs are the same (the prompt, no documents, no tools), the models reach for the same consensus answer. Their training corpora overlap on what counts as a sophisticated take.

Image models are less diverse than text models on a question with no concrete subject. If a product wants a surprising visual, asking "surprise me" is the wrong framing. A specific brief would move the 10 of 12 outputs away from the editorial-cartoon default.

The experiment is a template: 60 calls, hand grouping, one prompt, a published table. Any team can run its own version on its own question in an afternoon.

Source

Every AI model has the same "controversial" take on humans at ModelCompare. The Hacker News thread was at 205 points when checked.