If anyone, anywhere builds a superhuman artificial intelligence using present methods, the most likely outcome is catastrophe. There have accordingly been widespread calls for an international agreement prohibiting the development of superintelligence. In November 2025, MIRI’s Technical Governance Team published an example of one such agreement.
This post is an informal and very abbreviated summary of “An International Agreement to Prevent the Premature Creation of Artificial Superintelligence.” The full report is over fifty pages including appendices; for brevity, we necessarily omit a lot of detail. If you find yourself with questions about a particular subsection, it’s worth checking out the full report.
We will first discuss the Agreement’s context and purpose, then cover the highlights of its fifteen Articles, then explore a possible path to adoption.
Context and purpose of the Agreement
The race towards superintelligence poses a coordination problem. It is in no one’s interest to build superhuman machines which subsequently go rogue. But there are real benefits to AI capabilities, and states don’t wish to fall behind their rivals. What can be done?
While the technology is unprecedented, the international community knows how to solve this kind of problem. Like the nuclear nonproliferation agreements before it, the draft Agreement aims to prevent the proliferation of technologies that would be catastrophic for humanity if allowed to proceed unchecked.1 Specifically, it bars signatories from developing AI capabilities and technologies that may lead to superintelligence (Article I).
How might we formalize this restriction? We focus first on the large-scale training runs that currently produce frontier AI, setting thresholds beyond which training is monitored or prohibited (Articles II and IV). This alone would be a solid start, but large training runs are not the only source of rapid gains in AI capability. Over time, advancements in AI algorithms and hardware lower the bar to train a potentially lethal AI. Therefore, we also propose to restrict research in certain sensitive areas (Article VIII).
As the U.S. and China currently lead the world in AI capabilities by a substantial margin, this draft is framed as a bilateral agreement between the two countries, to be later joined by other nations. Authority to modify the Agreement and perform independent verification rests with the members of an Executive Council, initially and perhaps ultimately consisting of the U.S. and China. To help these states coordinate, we create a Coalition Technical Body (CTB) (Article III). The CTB would update key thresholds and standards to keep pace with new developments in the field of AI, and might also take steps to evaluate the capabilities of new AI models (Article XIII).
Software is considerably harder to monitor and regulate than nukes. We thus propose to control AI development primarily by controlling the hardware used to train AIs, which passes through several important bottlenecks. For instance, most AI chips are designed by NVIDIA and fabricated by TSMC on machines made exclusively by ASML.2 The draft Agreement thus obliges states to consolidate and monitor sufficiently large clusters of AI chips (Article V), to monitor the supply chain that produces them (Article VI), and to facilitate verification of how they are used (Article VII). It also obliges signatories to take domestic measures to enable verification of research restrictions (Article IX), coordinated by the Coalition Technical Body (Article X).
The creation of superintelligent AI anywhere poses a threat to the national security of every state on Earth. States which realize the gravity of this threat may be moved to act preemptively by taking military action against other states or actors attempting to build ASI. This draft Agreement seeks to avert military conflict by supplying feasible alternatives, including methods of verification in low-trust environments, a process for amending the Agreement (Article XIV), a means to resolve disputes (Article XI), and a process for orderly withdrawal from the Agreement (Article XV). It also codifies steps that states might take to prevent ASI development by others, including sanctions and asset restrictions. If other methods are ineffective, the Agreement acknowledges the right of states to conduct narrow, targeted cyber and military operations against facilities developing ASI (Article XII).
The Agreement in brief
The Preamble to the Agreement states that its primary purpose is to stably and verifiably reduce the risk of human extinction from artificial superintelligence. To that end, fifteen articles are laid out.
Article I prohibits signatories from developing or deploying artificial superintelligence and from permitting others within their jurisdiction to work towards that end. It does not oblige signatories to enforce this prohibition on non-signatories, but it does require them to “assist, or not impede, reasonable measures” to do so. This statement of the treaty’s primary purpose is modeled on Article I of the Non-Proliferation Treaty (NPT).
Article II defines the terms used in the treaty. The current draft sets training thresholds based on the amount of floating-point operations (FLOP) used in training: Runs above 1022 FLOP require monitoring, post-training above 1023 FLOP is prohibited, and any training above 1024 FLOP is prohibited.3 It also defines units of computing capacity by comparison to the common AI-specialized NVIDIA H100 chip, itself 3-10x the capacity of most consumer PC hardware. The threshold for compute clusters covered by the treaty is set at 16 “H100-equivalents.”4 The choice of computing thresholds is designed to prevent the training of superintelligence while excluding most household and personal use.
Article III outlines the structures that serve the coalition of signatories. An Executive Council, initially the U.S. and China, performs governance functions by unanimous consensus; it approves challenge inspections, sets budgets and policies, oversees the Coalition Technical Body (CTB) and appoints its Director-General. Individual CTB divisions track and verify chip use, consolidate info, evaluate models, and maintain manufacturing standards and research controls. The CTB also updates technical definitions, like FLOP thresholds, as needed, and the Director-General can make emergency updates to key thresholds that last up to thirty days before requiring Executive Council approval. Executive Council members can delegate tasks to the CTB or veto its decisions.
Article IV formalizes the AI training rules; each signatory agrees not to conduct training above the prohibited thresholds in Article II, nor permit such training within its jurisdiction. Any training runs above the Monitored Threshold in a signatory’s jurisdiction require a report to the CTB including code, data, and estimated FLOP. Trainers must allow the CTB logged and supervised access to data, with protections in place for security (e.g. CTB staff reviewing the data in person at the secure training facility). Such runs still require approval from both U.S. and Chinese inspectors onsite. The CTB then monitors the run, optionally requesting model data at specific checkpoints, and can pause the run if alarmed by its findings. It may also authorize certain kinds of specialized training runs, such as for narrow safety or medical purposes.
Article V obliges signatories to consolidate clusters of chips greater than 16 H100 equivalents (an amount of hardware that costs on the rough order of $500,000 as of 2025) into monitored facilities subject to CTB oversight. Facilities must be accessible and few enough to allow monitoring. The consolidation process itself is coordinated by the CTB and includes various measures to prevent cheating, such as the information sharing and challenge inspections covered later in Article X. The transfer, decommissioning, and destruction of chips also involve oversight. This article takes inspiration from START I for consolidation, inspection, and information sharing measures.
Article VI requires monitoring of the chip supply chain. AI chip production is already heavily bottlenecked to a few facilities; these facilities and their outputs would be subject to CTB monitoring or else shut down. Sale and transfer of advanced AI chips and chip manufacturing equipment would be, at minimum, tracked by the CTB and subject to CTB approval. This Article takes inspiration from IAEA monitoring and standards-setting for nuclear reactor facilities, from the INF treaty on inspection and limited perimeter monitoring, and from several arms control agreements on sale and transfer rules.
Article VII obliges signatories to permit chip use verification at monitored facilities. Facilities, under CTB supervision, must be able to confirm that chips do not perform prohibited training or research, and if the chips cannot be verified, the facility must be shut down. The CTB may choose to restrict metrics such as interconnect bandwidth that characterize prohibited AI training workloads. The CTB also coordinates fit-for-purpose security standards at monitored facilities and leads research and development of better verification methods. This Article draws on START I sharing of missile test telemetry, and the IAEA’s growing 24/7 camera and sensor presence at nuclear facilities.
Article VIII restricts research into AI algorithms and hardware that either advance general AI capabilities towards superintelligence or undermine verification efforts. It explicitly permits research that lacks these properties, and permits narrow applications such as medical diagnosis, scientific discovery, and industrial automation. It defines areas of restricted research to include distributed or decentralized AI training, more performant or efficient AI chips, fabrication methods of AI chips or components, and “methods for training general-purpose AI systems that would significantly increase model capabilities toward superintelligent performance or dramatically reduce the computational resources required to develop such systems.” Research meeting these criteria might be controlled (monitored and reviewed by the CTB) or prohibited (which no signatory may conduct, assist, encourage, or share, nor permit within its jurisdiction). Representatives from each signatory staff the CTB’s Research Controls division, which owns and updates these restrictions. Cited precedents include the “born secret” doctrine of the Atomic Energy Act of 1946, the decision-making duties of the National Nuclear Security Administration, and several domestic technology export regulations.
Article IX outlines mechanisms the CTB may implement to verify compliance with research restrictions, and directs each signatory to create or empower a domestic agency to enforce and document compliance within its jurisdiction. Precedents include the Enforcement chapter (18) of the 1946 Atomic Energy Act, sensitive data procedures in the U.S. Department of Energy, and the International Science and Technology Center, which gave Soviet nuclear researchers a peaceful alternative to weapons development.
Article X enumerates the duties of the Information Consolidation division of the CTB, established in Article III. It outlines the responsibilities of signatories to communicate relevant findings to this division and the channels available to them, and the responsibility of the CTB to protect its information from unauthorized access and conceal the identity of anonymous sources. It encourages, but does not oblige, signatories to contribute to verification efforts; it does oblige signatories to refrain from interference. It establishes protections and secure, confidential reporting channels for whistleblowers. It empowers signatories to request challenge inspections, or short-notice neutral examinations of covered facilities, and lays out procedures and limits on such inspections. Finally, it commits the U.S. and PRC to operating independent verification efforts using their own intelligence capabilities. Challenge inspections follow the example set by Part IX of the CWC; involvement of national intelligence in verification borrows from several nuclear arms agreements; and the CTB’s role as neutral information consolidator and its responsibility to secure and protect confidential communications follow the example of the IAEA.
Article XI sets the procedures and deadlines for resolving disputes related to the Agreement, giving signatories five days to respond to an inquiry. The Executive Council, with support from the CTB, may serve as a neutral arbitrator in such cases. If it confirms a violation, the Executive Council might require additional monitoring or restrictions on AI activities, require relinquishment of AI hardware, call for sanctions, or recommend that states take Protective Actions as defined in the next Article. Dispute resolution is modeled on provisions in the CWC.
Article XII authorizes signatories to take action against a state or actor given credible evidence of attempts to develop ASI. These actions might include trade and asset restrictions, visa bans, appeals to the UN Security Council, or, in extremis, targeted cyber or military action. In the latter case, Article XII obliges signatories to minimize collateral harm, to clearly and publicly state their reasoning, to terminate action once the targeted activities cease, and to refrain from employing force except in emergencies or after exhausting the alternatives. Cyber and military actions are subject to review by the Executive Council to determine whether they were necessary, proportionate, and properly targeted in accordance with Article XII. That states act to eliminate threats to their security is a reality predating any modern treaty, but precedents for explicitly legal military action nonetheless exist, e.g. in Chapter VII of the United Nations Charter.
Article XIII permits the CTB to securely test new AI models for the purposes of updating the training thresholds in Article IV. Summaries of test results may be shared, but details are secret by default, only released if the CTB’s Director-General deems it necessary to reduce the chance of human extinction from advanced AI. Precedents include the IAEA’s regular and detailed reports on developments in nuclear science and the implications for global security.
Article XIV describes the process for revising the Agreement. Subject to Executive Council veto, the CTB may update most definitions and implementation methods. Any fundamental changes require an Amendment, approved by consensus among the Executive Council. The Agreement is subject to a review conference every three years. Both amendments and periodic review conferences are borrowed from existing agreements such as the NPT and CWC.
Article XV outlines a process for withdrawing from the Agreement. Withdrawal requires 12 months of advance notice, during which the CTB and Executive Council may take steps to ensure the withdrawing state does not retain the capacity to pursue ASI, and the state is obliged to cooperate.
A plausible path to adoption
A common objection to proposals like this one is that states lack the political will to enact them. Historically, this is often true of many policies right up until it suddenly isn’t. Many plausible developments might prompt rapid action: an alarming crisis or near miss, a groundswell of popular support for AI regulation, or a sufficient appreciation for AI’s dangers among policymakers, for example.
Still, we don’t particularly expect the international community to leap directly to this draft Agreement from a standing start. Below is one way in which the nations of the world, especially the U.S. and China, might work their way to a feasible deal and navigate the subsequent period.
In the first stage, the U.S. and PRC issue a joint common-sense commitment not to use AI for purposes such as cyber attacks against critical infrastructure. Both countries begin domestically tracking AI hardware, chips, and research, and agree to limited transparency and information sharing. Early R&D begins on verification methods, as does diplomacy aimed at building international consensus on AI governance.
Next, the U.S. and PRC establish secure communication channels between high-level AI and cybersecurity officials. They periodically disclose information about the capabilities and hardware of high-consequence AI systems in their jurisdictions, and pre-notify one another of major training runs. They develop voluntary commitments and disclosures and continue to invest in safety and verification research.
Both countries then begin negotiating structured commitments and verifiable limits on training, and jointly test verification protocols. These commitments are firmed up into an international agreement restricting AI training and research that may lead to superintelligence or undermine verification. Additional countries join the Agreement, and the international community seeks to persuade the holdouts.
Ultimately, the Agreement buys the world breathing room to strengthen global coordination, build resilience into critical infrastructure, and perform urgently needed research into the safe design of superhuman minds.
Footnotes
- The full proposal also includes notes and precedents for each Article. Sources of precedent include the IAEA and its 1956 Statute; the 1970 Treaty on the Non-Proliferation of Nuclear Weapons (NPT); the 1991 and subsequent Strategic Arms Reduction agreements START I, START II, and New START; the 1983 Convention on Certain Conventional Weapons (CCW); the 1974 Threshold Test Ban Treaty; the 1922 Washington (Naval) Treaty; the Atomic Energy Act of 1946; the 1987 Intermediate Range Nuclear Forces (INF) Treaty; and several other historical agreements.
- There are also bottlenecks in chip components, such as high-bandwidth memory.
- For reference, as of 2025 it was estimated that about 30 models had been trained using more than 1025 FLOP, with the first such model being GPT-4.
- This is about the size of a compute cluster at a well-funded university lab.