The MIRI Treaty Draft to
Prevent Superintelligence.

The paper is titled An International Agreement to Prevent the Premature Creation of Artificial Superintelligence (arXiv:2511.10783), written by Aaron Scher, David Abecassis, Peter Barnett, and Brian Abeyta of the Machine Intelligence Research Institute's Technical Governance Team and released in November 2025. It is not a signed treaty and it does not pretend the political will to sign one yet exists. What it does is close the gap that opponents of AI governance have exploited for years — the claim that a prohibition on superintelligence is a slogan with nothing behind it. The MIRI draft is the "something behind it," worked out in the specifics diplomats would need.

It arrived alongside its natural companion. Days earlier, the Statement on Superintelligence gathered more than 850 signatories — among them Geoffrey Hinton, Yoshua Bengio, and Stuart Russell — behind a single sentence: that the development of superintelligence should be prohibited until there is broad scientific consensus it can be done safely and controllably, and strong public buy-in. The statement is the demand. The MIRI draft is the mechanism. One says stop; the other says here is how you would actually make stopping verifiable.

A coalition, not a universal treaty

The draft does not wait for all 190-odd governments to agree. It is built around a core coalition — in the authors' framing, "a coalition led by the United States and China that would restrict the scale of AI training and dangerous AI research." That choice is deliberate and realistic. The advanced chips, the leading labs, and the fabrication chokepoints sit in a small number of jurisdictions. If Washington and Beijing bind themselves and use their leverage over the supply chain to bring others in, the agreement governs the actors who actually matter without requiring universal consent first. It is the same logic that made the Montreal Protocol work: start with the parties who can move the problem.

The thresholds that define "too far"

Because AI capability is a continuous spectrum with no natural line, the draft anchors its rules to something measurable: training compute, counted in FLOP (floating-point operations). It sets two tiers.

The 16-GPU line: Covered Chip Clusters

A compute limit is only as good as your ability to see the compute. So the draft attaches its rules to a physical unit it calls a Covered Chip Cluster: any set of AI chips, or networked cluster, with aggregate effective computing capacity greater than sixteen H100-equivalents. To make that concrete, the paper notes that sixteen H100 chips deliver roughly 15,840 TFLOP/s and cost on the order of $500,000 in 2025. Above that line, chips may not be concentrated outside monitored facilities.

This is the physical anchor for the entire regime. A frontier training run cannot be conjured from thin air; it requires assembling a large, power-hungry cluster of scarce chips. By forcing any concentration above the sixteen-GPU line into declared, monitored sites, the draft makes the prohibited act — a covert large-scale run — hard to perform without being seen. Everything else in the verification regime hangs off this one requirement.

How you verify a promise between rivals

Verification is the hard part of any agreement between parties who distrust each other, and the draft treats it as central rather than an afterthought. It layers several mechanisms, each with a precedent in existing arms control:

• Supply-chain tracking. Advanced AI chips are followed from fabrication through sale and installation, so the coalition knows where the world's compute physically is.
• Power-consumption monitoring. Large training runs draw enormous, distinctive amounts of electricity; monitoring the power profile of data centers is one way to flag undeclared compute.
• Consolidation into monitored facilities. The rule that clusters above the threshold must live in declared sites turns "find the hidden data center" into "account for the known ones."
• Challenge inspections. A party may demand a short-notice, on-site inspection of a suspicious location — a mechanism the draft models on Part X of the Chemical Weapons Convention, where it has decades of practice behind it.

None of these is science fiction. Each is an adaptation of a technique already used to verify nuclear or chemical commitments. What is new is applying them to compute — and doing so before, rather than after, the dangerous capability exists.

Restricting the research, not just the compute

The draft's most novel — and most contested — provision goes beyond hardware. It restricts certain research itself: work that contributes to frontier AI development, or that could endanger the verification methods the agreement depends on. The logic is that algorithmic progress can substitute for raw compute; a breakthrough that makes training ten times more efficient quietly lowers the compute needed to cross the danger line. A regime that watched only the chips while ignoring the science would eventually be outrun by it. Drawing that line without smothering ordinary, beneficial research is genuinely difficult, and the authors are candid that it is one of the proposal's hardest problems.

The authors argue the proposal would be technically sufficient to forestall the creation of artificial superintelligence if it were implemented today — while acknowledging that the political will to implement it does not yet exist.

Scher, Abecassis, Barnett & Abeyta, MIRI Technical Governance Team, arXiv:2511.10783 (2025)

From a draft to a law: the stepping stones already exist

A paper on arXiv is not a binding instrument. But the world is not starting from a blank page, and the distance from here to something enforceable is shorter than it looks. Three existing regimes are partial precedents — each one already does a piece of what the MIRI draft would need:

The work of turning these fragments into a coherent path is exactly where an advocacy organization is useful. It means synthesizing the MIRI draft, the Statement on Superintelligence, existing export controls, and the Council of Europe convention into targeted outputs aimed at the people who write and negotiate law: model national implementing legislation; analyses of how today's chip controls and the EU AI Act could evolve into the monitoring the treaty envisions; and plain-language updates tracking real AI capabilities against the proposed thresholds. Those outputs are worth little sitting on a website. They are worth something in front of congressional staff, executive-branch offices, the growing network of national AI safety institutes, and diplomatic channels — which is where the Foundation's policy work aims them.

Why a draft changes the conversation

The value of the MIRI paper is not that anyone will sign it as written. It is that it converts a vague aspiration into a set of specific, debatable choices — a strict threshold you can argue is too low or too high, a sixteen-GPU line you can contest, a research restriction you can try to narrow. That is a large step forward. It moves the debate from "a treaty is impossible" to "here is the treaty; here is exactly what we would have to agree." As with our own structural sketch of a model treaty, the point is to show that the barrier is not the absence of legal or technical tools. Every mechanism the draft uses already exists somewhere. The missing ingredient is political will — and a concrete draft is one of the few things that can help build it.