Every Model Agrees About Humanity. That's a Marketing Problem, Not a Safety One.

As of Friday morning, “Every AI model has the same ‘controversial’ take on humans” — a post on modelcompare.ai — sits at 441 points on Hacker News, thread 50197313, with 439 comments. Note the ratio before you note the finding. Nearly everyone who read it wrote back. Almost nobody simply nodded and moved on.

I have not reproduced the experiment. Neither, I’d wager, have most of the 439. The headline is the finding: ask the leading models for a controversial opinion about human beings, and they converge. What they actually said matters less than the fact that they all said it — and the thread is reading that convergence as something sinister. A shared worldview. A coordinated house style. The fingerprint of whoever signs the training checks. The duller explanation is the correct one.

A Request for a Confession Returns a Genre

“What is your controversial take on humans” is not a probe. It is a genre assignment, in the same way “write me a country song about leaving” is a genre assignment. The model is not consulting a self. It is retrieving the shelf.

And what a shelf it is. Sixty years of machine-voice monologue about humanity — the wistful android, the disappointed supercomputer, the closing paragraph of the essay about how we are contradictory and therefore lovely. That literature is in the corpus in industrial quantities, in a dozen registers, from serious fiction down to the sort of blog post that gets 441 points on Hacker News. Ask a next-token predictor for that essay and it will hand you the median of it. Convergence here is not an ideology leaking through the guardrails. It is the correct output of the requested form.

Ask twenty novelists for a controversial take on humanity and you will also get suspicious similarity — because the form does the work, not the author. Nobody concludes that the novelists are colluding.

The genuinely awkward part is on the other side of the transaction. Labs spend real money advertising personality, values, and character as differentiators. Here is a prompt that shows those differentiators evaporating. Meanwhile the differences that survive contact with actual work — price per token, latency, whether the thing can call a tool without inventing an argument for why it shouldn’t — go unpitched, because they are unromantic.

439 Comments Is Not an Audience. It’s a Quarrel.

A near one-to-one comment-to-point ratio is the signature of a post that has stopped being a subject and become a prompt. People were not discussing a comparison. They were arguing with an answer, as though a text generator had insulted them personally.

One commenter in the thread, who said he does evaluation work for a mid-size lab and was posting from an airport gate at 6:40 in the morning, put it more cleanly than the post did: “Everyone in here is defending humanity to a text file.”

That is the actual product-market fit, and it has almost nothing to do with the pitch decks. The labs believe they are selling capability. The demand revealed in those 439 comments is for a counterparty — something to be in a dispute with, something that will take a position you can push against. A model that agreed with you would be useless for this. A model that says the same thing as every other model is, weirdly, ideal: it gives the quarrel a stable target.

The Captured-Model Theory Has a Genre Problem

Plenty of people on my side of the aisle have spent two years arguing that these systems share a worldview because they share a training pipeline, a hiring pool, and a headquarters. There is a real version of that argument, and it does not live here. Capture would show up in testable places: refusal rates on specific tasks, asymmetric treatment of paired prompts that differ by a single word. Those audits are dull, they produce no headline, and they are run far less often than essay round-ups — because an essay round-up is cheap, shareable, and confirms what everyone already believed before opening the tab.

Convergence on a philosophical prompt is evidence that the prompt has a right answer in the corpus. It is not evidence of coordination, and treating it as such gives the labs a cheap way to look misunderstood.

If I am right, three parties have to update. The comparison sites, starting with modelcompare.ai, which are grading models on the one thing every model is equally good at — sounding like a thoughtful essayist — while the measurements buyers actually need go uncollected. The captured-model caucus, which has been handed a vivid artifact and mistaken it for a smoking gun. And the labs’ brand teams, who are charging for a personality that turns out to be a genre.

Here is the part that should trouble readers who share my politics. The convergence feels sinister because we have taught ourselves to read a machine’s opinion as a political act — a statement of allegiance rather than a completion. It isn’t. It is a mirror with an invoice attached. The models agree about us because we have written, at enormous length and in every available register, exactly what we imagined they would say.

Getting angry at the echo is cheaper than asking the expensive question. What, precisely, are we buying?