The Antarctic Treaty is easy to overlook because Antarctica seems marginal. That is exactly why it is instructive. Here was a case where rival superpowers had overlapping claims to a strategically relevant space, every incentive to militarise it, and deep mutual suspicion — and they chose instead to agree that no one would press their advantage. They did not resolve the underlying dispute. They agreed to hold it in abeyance and to verify that everyone was holding to the deal. For a technology like AI, where the prize is contested and no one trusts a rival to hold back, that move — agreeing to not-yet rather than to a final settlement — is worth understanding.
What the treaty actually established
- Demilitarisation. Antarctica may be used for peaceful purposes only. Military bases, weapons testing, and manoeuvres are prohibited — the first arms-control agreement of the Cold War, signed before the Partial Test Ban Treaty.
- A freeze on claims. Seven countries had asserted territorial claims, some overlapping. The treaty froze them all: no claim is renounced, but none can be enlarged or newly asserted while the treaty is in force. The dispute is parked, not settled.
- Freedom of inspection. Any party may inspect any station, installation, or equipment of any other party, at any time, with observers having complete freedom of access. This is one of the most open verification regimes ever agreed.
- Science as the shared purpose. The continent is reserved for scientific research, with a duty to exchange results and personnel freely.
The transferable idea: freeze, then verify
The treaty's central innovation was to decouple the underlying rivalry from the immediate behaviour. Nobody had to concede who 'owned' Antarctica; they only had to agree that the question would not be pressed and that everyone could check. This is directly relevant to frontier AI, where the fear of ceding a decisive advantage drives the race. A treaty that demanded final agreement on who may build the most powerful systems would be unsignable. A treaty that froze the most dangerous forms of development — a moratorium on crossing specified capability lines — while guaranteeing mutual inspection, asks for something narrower and more achievable: not surrender, but a shared pause under verification.
The inspection regime is the part most worth copying. Antarctica's works because it is reciprocal and unconditional: the right to inspect is not a favour granted case by case but a standing entitlement written into the treaty. That symmetry is what made it politically acceptable to distrustful parties — each gave up secrecy in exchange for the same right over the others. Any AI verification regime that hopes to survive between rivals will need the same reciprocal, built-in quality rather than inspection by permission.
The Madrid Protocol: strengthening a treaty over time
Antarctica also demonstrates that a treaty can be reinforced as circumstances change. In 1991 the parties adopted the Madrid Protocol, banning mineral resource extraction and designating the continent a 'natural reserve devoted to peace and science', with the ban reviewable only after fifty years. A framework agreed in one era was hardened into a stronger prohibition in the next. AI governance is likely to follow the same arc: an initial agreement to freeze the most dangerous activity, strengthened later as understanding and trust accumulate.
Antarctica proves that rivals do not have to resolve who wins in order to agree that no one will make the dangerous move first. Sometimes the achievable agreement is not a settlement but a freeze — and a right to verify it.
Naoto Nakada, Founder · Nakada Foundation to Save Humanity
The limits of the analogy
Antarctica was governable partly because almost no one lived there and its resources were remote and unexploited. The stakes were low enough that restraint was cheap. Frontier AI is the opposite: the prize is immediate, enormous, and developed by private actors under competitive pressure, not just states planting flags. The cost of restraint is far higher, and the temptation to defect far greater.
But the mechanism does not depend on low stakes; it depends on mutual fear of the alternative. The powers froze Antarctica because an unrestrained scramble looked worse to everyone than a shared pause. That calculation — that a race no one can safely win is worse than a verified halt — is precisely the argument for AI restraint. The Antarctic Treaty shows the diplomatic form such a conclusion can take: not a grand resolution, but an agreement to stop short, together, and to let each other check.
Common questions.
A 1959 agreement, now with 56 parties, that demilitarised Antarctica, reserved it for peaceful scientific use, froze all territorial claims, and gave every party the right to inspect any other party's facilities. It was the first arms-control agreement of the Cold War and has held for over sixty years, later strengthened by the 1991 Madrid Protocol banning mineral extraction.
It shows that rival powers can agree to freeze a contested prize rather than settle who wins it, and to verify that freeze through reciprocal inspection. For AI, an agreement to halt the most dangerous forms of development — with mutual monitoring — may be more achievable than a treaty demanding final agreement on who may build the most powerful systems. The 'freeze, then verify' structure is the transferable idea.
It is reciprocal and unconditional. Any party can inspect any other party's stations and equipment at any time, with complete freedom of access, as a standing right written into the treaty rather than a permission granted case by case. That built-in symmetry is what made it acceptable to distrustful Cold War rivals, and it is the quality any durable AI verification regime would need.
Antarctica was cheap to leave alone — few people, remote resources, low immediate stakes — so restraint cost little. Frontier AI offers an immediate, enormous prize pursued by private companies under competitive pressure, so the cost of restraint and the temptation to defect are far higher. The mechanism still applies, but it depends on rivals concluding that an unwinnable race is worse than a verified halt.