A Draft of a Treaty, with Annotations

Article XV: Withdrawal and Duration

1. The Treaty shall be of unlimited duration.

2. Each Party shall in exercising its national sovereignty have the right to withdraw from the Treaty if it decides that extraordinary events, related to the subject matter of this Treaty, have jeopardized the supreme interests of its country. It shall give notice of such withdrawal to the ISIA 12 months in advance.

3. During this 12 month period, the withdrawing state shall cooperate with ISIA efforts to certify that after withdrawal, the withdrawing state will be unable to develop, train, post-train, or deploy dangerous AI systems, including ASI or systems above the Treaty thresholds. Withdrawing states acknowledge that such cooperation aids the ISIA and Parties in avoiding the use of Article XII.

   1. In particular, the withdrawing state, under ISIA oversight, will remove all covered chip clusters and ASI-enabling assets (e.g., advanced computer chip manufacturing equipment) from its territory to ISIA-approved control or render them permanently inoperable (as described in Article V).

4. Nothing in this Article limits the applicability of Article XII. A State that has withdrawn (and is therefore a non-Party) remains subject to Protective Actions if credible evidence indicates activities aimed at ASI development or deployment.

Precedent

It is common for treaties to lack expiration dates. The first paragraph of Article XVI of the CWC states “This Convention shall be of unlimited duration.”

Treaties of unlimited duration do not necessarily last forever.^[1] But they do typically provide a mechanism for withdrawal, usually with a required period of notice and other stipulations that might let it leave in a manner less concerning to the remaining parties. Article XVI of the CWC allows for a party to withdraw “if it decides that extraordinary events, related to the subject-matter of this Convention, have jeopardized the supreme interests of its Country.” The withdrawing country must give 90 days notice. Article XVI of the Outer Space Treaty requires one year notice for withdrawal.

Our treaty language expects 12 months notice from departees, allowing ample time for assisting with the assurance-providing measures in Paragraph 3. Our intent with these measures (which go beyond what we readily find in the historical record of withdrawal provisions) is to reduce the potential need for protective actions against the withdrawing Party, as no Party or non-Party can be allowed to create ASI or weaken the world’s ability to prevent its creation.

Historical precedent for a withdrawn party remaining subject to protective actions is found in the case of United Nations Security Council Resolution 1718, which imposed sanctions against North Korea after its 2006 nuclear test, despite North Korea’s previous withdrawal from NPT.

Notes

Given the dangers of ASI research and development, as well as the risk that if one country decides to withdraw from the treaty and race to superintelligence then others might follow, a treaty needs barriers to withdrawal.

In practice, this is challenging. North Korea, for example, withdrew from the NPT to continue its nuclear proliferation activities, even at the cost of UN Security Council resolutions and associated sanctions. The consequences did not prove sufficient to cause North Korea to stop its proliferation activities.

If nations wish to withdraw from the treaty, our wording makes it clear that, in the eyes of all parties, they forgo the right to AI infrastructure, and that they would be subject to Article XII Protective Actions if they engage in dangerous AI activities. Any further negotiation around the ASI issue — e.g., to avoid Protective Actions — would have to be negotiated separately by interested parties.

Parties concerned about withdrawals could include mechanisms to make withdrawal more difficult. For example, both U.S. and Chinese officials could agree to install mutual killswitches inside retained datacenters, allowing either party to permanently shut off the other’s datacenter. Alternatively, parties to the treaty could adopt a multilateral licensing regime in which all new AI chips must be fabricated with hardware locks that require approval from multiple parties to continue operation, so that if a country withdrew from the treaty, others could stop approving their licenses and incapacitate their chips. Another option involves moving key AI infrastructure into third-party countries where the infrastructure could be confiscated or destroyed if a party withdrew from the treaty. Our draft sticks to minimal deterrence methods, but many other methods are available (or could be made available with a little technological investment).

Our draft treaty is focused narrowly on preventing the creation of rogue artificial superintelligence. A more sweeping treaty could also attempt to unify the parties around a particular positive vision of how AI development could eventually continue, and could end with an article that agrees to joint investment in that vision. Coming to that sort of agreement seems to us like an extra difficult step, and so our draft does not venture that sort of proposal. As we discussed elsewhere, people need not agree on those details to agree that the race to superintelligence should be stopped, and world leaders could unite around a treaty that resembles this draft even as they separately negotiate joint investments into positive paths forwards to the future, in whatever ways they see fit.