The Global AI Accord

The atmosphere was saved by agreement. The cognitive commons can be too.

An independent, non-commercial consultation text proposing binding international governance of artificial intelligence, modelled on the Montreal Protocol. It seeks no signatures. It seeks authors.

Read the full Accord (PDF, Version 1.0)

Status

The Global AI Accord is an independent, non-commercial consultation text.

The Declaration

A Declaration on the Governance of Artificial Intelligence as a Global Public Interest

Issued for global consultation

Every generation inherits the consequences of technologies it did not choose. Ours is the first to create a technology that may choose for itself.

Artificial intelligence is not the first invention to change how humanity thinks. The alphabet did that, and the printing press, and the broadcast tower. It is the first that can do our thinking without us: reading, writing, judging, deciding and acting at a scale and speed no human mind can follow. Where every earlier technology extended the human hand, this one substitutes for the human mind, and its effects do not stop at any border.

This Declaration is not written in fear of that power. Within a single decade, a learning machine solved a problem that had defeated biology for fifty years, the folded structure of proteins, and the answers were given freely to every scientist on Earth. Hundreds of millions of structures, the raw material of new medicines, placed in the commons by choice. That act proved what no argument could: the fruits of machine intelligence can be shared as the inheritance of all humanity, when those who hold them decide to share.

And yet the same decade that gave science this gift has set algorithms to hunt human beings on the battlefield, and begun to thread machine judgement into the systems that watch for nuclear attack. The question before the world is not whether this technology is good or evil. It is which of these two futures we will organise ourselves to choose, and whether we will choose at all, or let the race choose for us.

The code may be written in one country. The data centres may stand on one nation's soil. But the consequences of autonomous intelligence, once released, belong to everyone. A system that erodes truth erodes it everywhere. A system that escapes control escapes it everywhere. A weapon that kills without human judgement diminishes the humanity of us all.

For this reason, we declare that artificial intelligence must be understood as belonging to the category of the global commons, alongside the atmosphere, the oceans, the polar continent and outer space: domains that no nation may own, that every nation affects, and that only all nations together can protect.

Recalling that in 1987, in Montreal, the nations of the world confronted a class of industrial chemicals that were profitable, embedded in the global economy, and destroying the layer of the atmosphere that shields all life, and that they chose, before the damage was complete, to bind themselves to a common discipline;

Recalling that the Montreal Protocol succeeded not through goodwill alone but through design: it acted on scientific evidence before the harm was irreversible, it tightened its own restrictions as knowledge advanced, it helped poorer nations to comply rather than punishing them, and it made compliance measurable, so that every party could verify the conduct of every other;

Recalling further that rivals and adversaries have repeatedly bound themselves to one another when the alternative was mutual ruin: in the Antarctic Treaty signed at the height of the Cold War, in the direct line agreed after the Cuban crisis, in the conventions renouncing biological and chemical weapons, and in Montreal itself, which the great powers of a divided world joined together;

Recognising that the development of frontier artificial intelligence is presently governed by a race: a race between companies, in which safety yields to speed of release, and a race between states, in which caution is mistaken for weakness. These are not two problems but one, the same competitive logic at two scales, and no participant can escape it alone. A race without rules is not won by anyone. It is only survived, or not;

Recognising that no nation, alliance or bloc can secure the safety of artificial intelligence by excluding its rivals. Intelligence encoded in software cannot be fenced. Export controls delay; they do not prevent. Every strategy of exclusion invites the excluded to build in parallel, in secret, and without restraint. The governance of a borderless technology by a coalition of the willing is a contradiction in terms. It requires a coalition of the whole;

Now, therefore, this Declaration affirms the following principles:

The red line. The authority to take a human life must never be delegated to an algorithmic process. In every use of lethal force there must remain a human being who decides, who can refuse, and who answers for the decision. Autonomous weapons that select and engage human targets without meaningful human control must be prohibited absolutely, and all other autonomous weapons must be subject to binding regulation. No algorithmic process may ever hold authority over nuclear weapons, and no system of early warning may be permitted to compress, beyond the reach of human deliberation, the decision on which the survival of civilisation depends. More than once in living memory, the world was saved from nuclear war by a single human being who doubted what the machines were telling him. To remove that human is to remove the only safeguard that has ever actually worked.

The cognitive commons. The integrity of the shared information environment is a common inheritance of humankind. No state and no enterprise may treat the deliberate, automated degradation of human understanding as a legitimate instrument of competition or conflict. What a people holds to be true is the ground on which every one of its other freedoms stands, and to poison that ground by machine is an act against all humanity, whoever the intended target.

The common registry. The physical foundations of frontier artificial intelligence, the fabrication of advanced semiconductors and the operation of computing infrastructure at the largest scales, shall be subject to international transparency, so that the concentration of capability is visible to all parties, as the production of ozone-depleting substances was made visible under Montreal.

The shared dividend. The benefits of artificial intelligence in medicine, in science, in agriculture and in education shall be made available across all nations, through mechanisms modelled on the multilateral fund that allowed developing countries to join the Montreal Protocol as partners rather than supplicants. The gift of the protein structures showed that this is possible. A world divided into cognitive empires and cognitive colonies will not remain at peace.

Verification and revision. An agreement that cannot be verified is an aspiration, not an accord. Compliance with these principles must be demonstrable through evidence, inspection and independent audit, and the accord itself must be built, as Montreal was built, to tighten and adapt as the technology advances, without requiring the world to negotiate its survival anew each decade.

The precedent is imperfect. The chemicals of 1987 were made by a handful of firms and had substitutes waiting; intelligence is diffuse and has none. But the lesson of Montreal was never chemical. It was political: sovereign nations, in full knowledge of the profits at stake, chose the survivable future over the convenient one, and held each other to that choice through verification rather than trust.

This Declaration is addressed to no company and against no country. It is issued in the name of no institution and in the service of no market. It asks a single question of every government, every scientist and every citizen: whether the wisdom of our politics can be made to match the reach of our technology, before the answer is chosen for us.

The atmosphere was saved by agreement. The cognitive commons can be too.

This text is issued for consultation, and consultation means exactly what it says. To every government that has argued the hour is not yet right: say so here, and say why. To every scientist who can find the flaw in these pages: find it. To every citizen who has felt these questions and been told they belong to experts: they belong to you. Take this text. Argue with it, correct it, translate it, carry it. It seeks no signatures. It seeks authors. The accord humanity needs will be written by many hands, or it will not endure at all.

Version 1.0. Issued for consultation. This document may be shared freely in unmodified form.

The Full Accord

The Declaration above is Part One of the Accord. Part Two, the Explanatory Memorandum, sets out the reasoning, the evidence and the precedent behind each principle in eight sections, and answers the strongest objections rather than the weakest: the cognitive commons, the Montreal precedent, the honest limits of the analogy, the failure of every alternative, the verification gap, the automated battlefield, the pillars of the Accord, and a covenant with tomorrow.

Download the complete text (PDF, 24 pages, Version 1.0)

Permanent archival record: https://doi.org/10.5281/zenodo.21204189

The Consultation

This text is unfinished on purpose. Version 2.0 will be shaped by what arrives, and substantive contributors will be acknowledged.

Write to: consultation@theglobalaiaccord.org

Every serious letter will be answered.

About This Text

The Global AI Accord is an independent, non-commercial work. It was produced with no funding, is issued in the name of no institution, and is unaffiliated with any company or product. It seeks no signatures and collects no data.

Initiated by Alexandra Carvalho, independent researcher (ORCID 0009-0003-0156-7180), and offered to the world to finish.

Licensed under Creative Commons Attribution-NoDerivatives 4.0 International (CC BY-ND 4.0). This document may be shared freely in unmodified form.

Consultation Text, Version 1.0, July 2026.