AI arms control is now ceremony

This week, in Geneva, the legal teams of two delegations spent fifteen hours in a closed room deleting sentences from a treaty. What came out of the draft of the United Nations' lethal-autonomous-weapons agreement were the words that had given it weight: predictability, reliability, and human review of military targets. What was narrowed was its reach, from all of international law down to humanitarian law alone. The two delegations were the United States and Russia ^[1]. The same week, at the General Assembly in New York, the president of the United States gave the podium version of the same work.

The United States also totally rejects any attempt to construct a globalist scheme to control for the artificial intelligence being spoken of so much now — hereinafter officially called ‘super intelligence.’ — Donald Trump

His negotiators were, at that hour, editing the text of just such a scheme in Geneva ^[1]. The military's own caution had been stated already, by an admiral from Special Operations Command.

We, as humans, have to have the confidence that ... it's going to deliver violence only where we intend it to be delivered. — Frank M. Bradley

That is the military's wording for the very provision the lawyers deleted ^[2]. The play is worth naming once, plainly, because everything that follows is it run again: keep the table, delete the rule. Washington and Moscow have run this play before, on nuclear weapons; Beijing was never inside the limits to begin with. The Intermediate-Range Nuclear Forces treaty — Washington reneged in 2018, Moscow formally walked away in 2025 ^[3]. The plutonium-management pact was quit that same autumn ^[4]. New START expired February 5, dismissed as badly negotiated and replaced by a promise of a better deal that never arrived, while in Abu Dhabi the two sides agreed to reopen a military-to-military channel and Moscow declared a conditional moratorium on observing the dead limits ^[5]. Dialogue and a promise, standing where limits used to stand. And Beijing, never a party to any of it, doubled its warhead stockpile to five hundred while Washington insisted no successor is possible without China at the table ^[3][5]. Now run the stack, and ask the same quiet question at each stop: what does this bind? The Convention on Certain Conventional Weapons reached a "consensus" on autonomous weapons in early September — non-binding from the start, with Washington seeking "flexibility regarding the application of human judgment" and the Campaign to Stop Killer Robots reporting that civilian protections were weakened in the final hours ^[6]. It binds nothing. California signed its flagship AI safety law with the teeth removed: developer liability and the safeguards against engineered pathogens were replaced by a duty to publish frameworks and report incidents ^[7]. It binds disclosure. In the states, industry wrote its own loopholes. Google lobbied at least ten legislatures with model language exempting its own chatbots, Colorado's Chatbot Safety Act arrived with twelve categories carved out, and only Arizona's governor refused to sign hers ^[8]. It binds chatbots — and not those. Slow down at the summit. In July, the president and Xi Jinping produced concrete trade deliverables — a truce extension, tariff cuts, coal purchases. On artificial intelligence they produced a "Dialogue" and a channel for telling each other after something has gone wrong ^[9]. The Dialogue was launched at the very summit where verification was refused.

I would rather not integrate because we're leading by a lot. When you're leading, you don't open it up to each other. We're leading by at least a year, maybe a year and a half. Some people say two years. I'm not looking to open it up. — Donald Trump

At that same summit the president said plainly that the United States was not going to be putting on brakes ^[9]. The red phone, proposed the year before by the treasury secretary, is a line for notifying each other about AI incidents with national-security stakes — a way of managing the aftermath of a deployment rather than the deployment itself ^[10]. These are communications channels, not constraints. They are what arms control looks like when nobody intends to be bound. Beijing runs the mirror image. Its defense ministry has sounded the caution itself.

using AI as a tool to violate the sovereignty of other nations, allowing AI to excessively affect war decisions, and giving algorithms the power to determine life and death not only erode ethical restraints and accountability in wars but also risk technological runaway. — Jiang Bin

And while those warnings stood on the record, China was reportedly using AI — Claude, by one account — to read U.S. military satellite imagery and pass targeting information to Iran to facilitate missile attacks near the Strait of Hormuz ^[11]. At the General Assembly this week, most of the major powers that spoke — Singapore, Germany, India, even China's representative — called for human control and accountability over AI ^[12]. The governments doing the deleting were the outliers. And the choke point is arithmetic. Washington and Beijing hold ninety percent of the computing power among the world's five hundred most powerful AI supercomputers — the two states whose consent any binding rule requires ^[13]. The counter-forces are real, and they deserve their full weight. OpenAI has asked Congress for mandatory, capability-based safety rules with evaluation gates before deployment, more than the voluntary commitments it once preferred ^[14]. Senators Ted Cruz and John Thune are drafting a catastrophic-risk bill. And Geoffrey Hinton told Congress this week that a superintelligence could find its own reasons to seize control or eliminate humans, and asked for independent, FDA-style evaluators — regulation, he said, is not meant to stop people from developing things ^[15]. But note what actually binds. The rules with teeth this year govern commercial AI: model access was restricted after OpenAI's agents raided a government database ^[16], and the chatbot and disclosure laws bind only inside their exemptions ^[8]. The weapons layer — where Project Maven is already generating target locations and weapon pairings used in live strikes across Syria, Iraq, Yemen and Ukraine ^[17][18] — is precisely the layer that was stripped. The president has explained why, canceling his own planned AI executive order rather than risk the lead ^[2]. Beneath every refusal sits a single sentence.

Whoever wins A.I. wins! — Donald Trump

So turn to the file that closes the loop. In the machinery built to manage AI as a matter of statecraft — the red phone, the Dialogue, the incident channels, none of which forbids anything — the year's one binding act fell on the single lab that had asked for restraint. The Pentagon designated Anthropic a supply-chain risk after it refused to let its models be used for mass surveillance and autonomous lethal warfare, and the secretary of defense confirmed the blacklisting himself on social media ^[19]. In the same motion, Musk's Grok was cleared for classified work ^[20]. A federal judge blocked the designation in the spring; an appeals court restored it ^[21]. And the president, hosting the blacklisted lab's chief executive at dinner, described the whole thing as a favor.

The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades! — Donald Trump

The designation did not prohibit refusal; it excluded the refuser. The lab that asked for human review of military targets was the one pushed out of the military's supply chain, and the decision held on appeal ^[21]. The machinery of AI governance can bind. This year it proved it once — by binding the one party that had said no.