A new paper names three ways the AI race can end badly. The third one is the new one.

The argument for building superintelligence quickly usually rests on a single fear: that somebody else gets there first. It is the logic of every arms race, and it has become the standard reply to anyone asking the labs to slow down.

A preprint posted to arXiv on October 12, 2026 argues that this framing is missing two thirds of the problem.

The paper is "Superintelligence and the Triple Risk of an AI Race," by Matthew Rendall, Max Hellrigel-Holderbaum and Leonard Dung. It has not been peer reviewed. It is a philosophy and policy argument rather than an empirical result, and it should be read as one.

The three risks

The authors argue that a push to reach artificial general intelligence first carries three distinct dangers, and that most debate collapses them into one.

First, being overtaken by a rival. This is the familiar one, and the one the race argument is built on. If a competitor reaches a decisive capability lead, the loser's position may be unrecoverable.

Second, triggering a war. Also familiar from the history of arms races, and usually left out of AI discussion entirely. A state that believes a rival is about to achieve a decisive advantage has an incentive to act before that happens. The danger is not the finished technology. It is the stretch of time just before it arrives.

Third, losing control to agents far more intelligent than we are. The authors treat this as the new risk, the one with no precedent in previous arms races, and they frame it in terms of human survival rather than competitive disadvantage.

The structure of that argument is the useful part. Risks one and two pull in opposite directions. Racing harder reduces the chance of being overtaken and raises the chance of provoking a conflict. Risk three is different again, because it does not care who wins. A race won carelessly produces it just as reliably as a race lost.

The claim that will get argued about

The paper also makes an empirical claim, and it is the one most likely to draw fire: that China does not currently appear to be racing toward superintelligence.

If that is right, a large part of the public case for American urgency rests on a race that only one participant has entered. The authors use it to argue the United States should put more weight on AI safety, both on its own and in cooperation with others.

If it is wrong, the argument weakens considerably. This is the part of the paper a reader should check against other sources rather than accept, and it is the part most sensitive to what counts as racing. A government that funds frontier labs heavily while saying little in public looks very different depending on which evidence you weigh.

Why the separation matters

There is a lot of writing about AI risk that is really writing about one risk. The useful move here is the separation, because the three dangers call for different responses.

If you are worried about being overtaken, you invest and you accelerate. If you are worried about provoking a war, you communicate, you build verification, and you avoid the appearance of a sudden breakout. If you are worried about losing control, you slow down at exactly the moments the first concern says to speed up.

Any policy that treats the AI race as one problem is going to pick one of those and quietly fail at the other two. Naming them separately is not a solution, but it does make the tradeoffs visible, which is more than most of this literature manages.

What it does not do

The paper does not provide a technical path to control, and does not claim to. It does not put numbers on any of the three risks. It is an argument about how to frame a decision, not a measurement of anything, and readers who want evidence on capability or timelines will not find it here.

It is also a preprint. It has not been through review, and the claim about China is the kind of assertion that review would press hardest on.