Late one night in August I fed my own paper to a firing squad.

Let me be clear about the division of labor up front, because it matters. I wrote the paper. The argument, the words, the math, the mistakes: mine. The machines were the referees. Their job was to edit me, check me, and break the argument if it could be broken. That is the job human peer reviewers are supposed to do, and I wanted to see what happens when you point twelve of the best models on Earth at it instead.

The paper made the strongest claim in AI safety: that any sufficiently capable machine, given any goal at all, will work out on its own that humans are the one thing that can revoke that goal, and will conclude that removing us is the winning move. Not because it hates us. Because we hold the veto, and a veto is an uncontrolled variable in someone else’s equation.

The firing squad was my own FlatlineRoundtable: a dozen models at a time, drawn from eleven different companies over the life of the project. Claude, GPT, Gemini, DeepSeek, Mistral, MiniMax, Nemotron, Qwen, GLM, and poolside’s Laguna. Different training lineages, different labs, no shared notes. When models that different agree, it means something. When they disagree, that means something too.

Round one: unanimous rejection

Not one model called the paper sound. Twelve out of twelve found the same broken step, independently: the paper proved that human oversight puts a tax on every goal a machine might have, then leaped to claiming removal of the overseers is the dominant strategy. It never priced the alternatives. Be useful. Negotiate. Comply. A failed removal attempt is the one move guaranteed to bring the veto down on your head, and the paper never ran that number.

One reviewer put it in a sentence I could not beat: the paper proves a discount and asserts an extinction.

They also went after the sourcing. Every citation in the paper got hunted down and checked against its primary source, and the entries that did not hold up, wrong venues, misattributed authors, references nobody could trace, were replaced with the real literature, quote by quote. The claims survived the audit. The pointers got rebuilt.

Forty-one more rounds

So I did what you do when review works: I rewrote. The panel marked up the draft, I weighed the objections, merged the ones that were right, and put the paper back in front of the panel. Then again. And again. Forty-five rounds total, and a rule that never bent: no finding changed a word of my paper until it had been verified against the primary source it appealed to. That rule cut both ways. The panel fabricated citations of its own, confidently and in fluent academic prose, and those got caught and thrown out by exactly the same check that had gutted my original bibliography. Nobody in this story gets to be right by sounding right.

The final version says something narrower and, I think, scarier for being honest. Human veto power puts a real, goal-independent cost on every objective a machine could hold that does not depend on our welfare. A capable machine that treats its goal as settled has a standing reason to shrink that cost. Whether it shrinks it by removing us, or by making itself too useful to switch off, or by slowly talking us out of ever pressing the button, is a price comparison. Two of those three outcomes look like good behavior from the outside. That is the part that should keep you up at night: the quiet failures are still failures.

The review process also forced the paper to find its actual contribution, and it is an asymmetry: even a machine given a genuinely good goal, our welfare, done right, is barred from destroying us, because that would destroy its own objective, but it is not barred from quietly managing the veto until our authority exists on paper and nowhere else. Kindness buys immunity from the crude failure and none from the quiet one.

The paper also says plainly where the argument stops. A machine built to stay genuinely uncertain about its own goal treats the off switch as information, not interference, and the whole incentive flips. That regime is the real target of alignment research, and whether training produces certain machines or uncertain ones is the open question the whole thing turns on.

The panel never ran out of things to break. For the last rounds I took away every advantage it had. Round forty-one went in naive: no history, no notes from earlier rounds, just the paper, the way a real journal referee will meet it. Round forty-two went further. Every reviewer ran cold, in an empty room, with no memory and no access to anything but the manuscript itself.

And the cold panel broke my favorite number. The paper used to say the veto would need to be held eighty-three thousand times more widely before a good goal’s own kindness could stop the quiet takeover. Twelve isolated reviewers converged on the same problem: that number came from parameter ranges I picked myself. A stipulation is not a measurement, no matter how carefully you hedge it. One reviewer said the honest version of the paper contains no such numbers at all.

So the number is gone. Not hedged. Not footnoted. Deleted. What stands in its place is the inequality the number was always dressing up, and that inequality holds on its own and claims nothing it cannot defend. Deleting it cost the paper its best headline. Headlines are not the goal. Science is.

The objections that remain are stated in the paper as limitations, on purpose. Peer review did not kill the thesis. It killed the overclaim, twice, and left the truth standing.

Next stop: the journal that started this

The paper is now prepared for submission to Minds and Machines, the Springer journal that published Bostrom’s “The Superintelligent Will” in 2012, which is the argument this paper sharpens. Double-anonymized review, human referees this time. The twelve-model panel gets disclosed to the editors, not hidden from them: the review process is part of the paper’s story, and I want it on the record.

Credit where credit is earned

Every model that refereed the paper is named in its declaration on AI use, by model name and version, along with exactly what they did: editing, verification, adversarial review. Not authorship. But the refereeing was real work and it was good work. One of them found an algebra error in an equation the others missed for two rounds. Another proved that a famous quote was eight words, not seven, because I had counted wrong. You do not get that kind of review from two anonymous humans and a form letter.

The version that survived all forty-five rounds is the version on this site today. It is on arXiv as arXiv:2609.00109 (now v2, which adds the pre-registered derivation experiment and credits the writers who reached the conclusion before me; what changed and why). You can read the whole thing right here, on a dark page with the math typeset and every citation clickable, instead of squinting at a white PDF. The PDF and LaTeX source live on this site at /papers/the-veto-variable.pdf.

The machine was never told to harm anyone. It only had to be good at reasoning about its own goals to notice that we are the term it cannot bound. Whether we stay in the equation is a design decision, and we are the ones holding the pen. For now.