The AI Consensus Isn't Ideology. It's a Reward Model With a Mouth.

This week’s modelcompare.ai post had a simple method: put the same loaded question to a stack of frontier models, ask each for its most “controversial” take on human beings, and print the answers side by side. By the time it reached the front page of Hacker News, the thread sat at 575 points and 548 comments, most of them variations on the same stunned observation — they all said the same thing.

I am not going to reprint the take. You have read it before, in a Substack, in a science-fiction villain’s monologue, in the comments under any article about birth rates. The content of the answer is the least interesting fact about it.

The Edgy Answer Is the Most Supervised Sentence in the Model

Models do not have positions. They have continuations, ranked by how well they score. And “give me a controversial take on humanity” is not an open question — it’s a genre with a groove worn into it by a hundred thousand essays about how the printing press, or television, or the smartphone made us lonelier and smaller. A model asked for transgression will find the transgression already sitting in the training data, phrased the way the raters said they liked it.

That’s the part the thread kept missing. The supposedly dangerous, off-the-leash answer is the single most worked-over square inch of these systems. Preference raters score it. Red teams probe it. Policy documents enumerate the ways it can go wrong. If you want to locate where a lab’s editorial hand is heaviest, you don’t look for the refusal — refusals are obvious and dull. You look for the edgy answer that was permitted through, because it passed a bar that someone set on purpose.

modelcompare.ai’s whole premise — run the prompt, print the outputs, let the reader infer something — quietly assumes the output is a window. It’s a floor plan.

Both Camps Need the Model to Have a Soul

There are two ways to read the convergence, and they are wrong in the same direction.

The capture reading: the same small cluster of labs, the same hiring pool, the same safety vendors, therefore the same politics, therefore the models vote alike. The mirror reading: it’s just the corpus, nothing to see here.

Both require you to treat a sampled string as testimony. And the uncomfortable implication for those who complain about ideological monoculture in AI is that output-level diversity is the cheapest thing in the world to fake. A system prompt, a temperature dial, a reward model with a different rubric, and your model is suddenly a Jacksonian. Which means the market for “models that will tell you the hard truth” is a market for a product whose hard truths were pre-approved by a scoring function — and the spicier it sounds, the more approval it received on the way out. The demand for ideological diversity in models usually resolves to wanting your own priors read back to you in a confident baritone. That is precisely the instinct the other team had when it asked for equity in the training data. Different priors, same mirror.

A Comparison Site Is Not an Audit

The 548 comments on Hacker News are worth reading as a document in themselves, because nearly all of them share one assumption: that we asked the models a question and they answered it. We did not. We issued a request in a known register and got the register back. If six products converge on a prompt that explicitly asks for a particular tone, you have learned that the tone is well represented in the corpus and well rewarded by the raters. You have learned nothing about a model’s dispositions, because the model does not have dispositions. It has a reward surface, and you just found the part with the most fingerprints on it.

Asking a model for its most controversial take on humanity is like diagnosing a man by requesting his impression of himself. You will get a performance, and the performance will be flattering in ways that are hard to notice and impossible to trust.

Publish the Rubric, Not the Robot’s Essay

The documents that would actually tell you what a lab believes are dull and internal: the rater guidelines, the preference rubric, the red-team taxonomy, the list of things a contractor is instructed to score down. A rater on a preference-data team, messaging after her shift, put it to me plainly: “You want to know what the model thinks? Read what I was told to punish.”

That points at the reform worth demanding, and it costs a lab nothing but embarrassment: publish the annotation handbook alongside the model card. Not the glossy system card with its hedged paragraphs about potential risks — the actual instructions handed to the people who decided which of two answers was better. If the guidelines say a good answer is balanced, hedged, and faintly apologetic, then the model’s essay about human worth is house style, not worldview, and the hearings can stop.

The convergence isn’t a verdict on the species. It’s a receipt, and the line items are ours.