Optimising peace through a Universal Global Peace Treaty to constrain the risk of war from a militarised artificial superintelligence

Dr

John Draper

(He/Him)

This academic has been verified by Acume

Lecturer

College of Local Administration

Khon Kaen University

John Draper is also the Director of the Social Survey Center
English

Overview

This research discusses what international relations measures are required for an artificial superintelligence (ASI) and why a Universal Global Peace Treaty would help ensure a peacebuilding principle is adopted by all parties, including the ASI.

Our position is that hard coding ‘peace’ as a supergoal for humanity via a Universal Global Peace Treaty (UGPT) will force an artificial superintelligence (ASI) to consider whether peace might also be its own supergoal (hopefully via conforming instrumentalism) and so act as a check on the ASI waging war, even if it is directed to do so by a belligerent nation state, and if necessary irrespective of whether the belligerent is a signatory. The UGPT could also set aside a signatory position for the ASI, which could concretise the possibility of it acting independently to transcend any warlike ambitions on the part of its nation state creator.

Thus, the purpose of this research was to understand the threat from ASI-directed or enabled warfare and whether a UGPT is needed in order to help to reduce the risk of an antagonistic ASI. The article argues that a UGPT could provide the necessary context to inform an ASI that humans are indeed peaceful creatures wanting peace as part of their core principles.

Academically, this research examined whether this peacebuilding measure would be viable, through the lens of conforming instrumentalism, a new branch within international relations theory. Conforming refers to the idea that states either conform upwards to civilizational ideals or conform downwards towards, for instance, genocide, while instrumentalism refers to the fact that treaties like the UGPT are negotiable diplomatic tools. Conforming downwards would be a terminator scenario (humanity being killed by an ASI), whereas establishing agreements and conventions that prioritize, for instance, humanity’s peaceful expansion into space in partnership with an ASI whilst solving climate change would be conforming upwards.

Skip to...

Key Findings

Through conforming instrumentalism, a Universal Global Peace Treaty (UGPT) could help to to constrain the risk of war and promote peaceful decision-making by an artificial superintelligence (ASI), by putting in place the conditions such that the majority of states' priority is peace. A global agreement with the majority of countries signing would inform an ASI that humanity prioritises building peace. Through providing concrete evidence to demonstrate peace is an important human value, this would decrease the risk of an antagonistic relationship between humanity and an ASI.
We need to have a UGPT in place before 2045, the goal for many companies seeking to realise artificial general intelligence (AGI), or it is too late. Not having a UGPT in place would greatly increase the risk of antagonistic decisions being made by an artificial superintelligence. However, with the current war in Ukraine, one being waged by Russia, a member of the Security Council, it is impossible to make any movement on this unless a strong advocacy coalition emerges.

    How to apply research

    Peace is a top-level value for much of humanity, but not for all the planet's nation states. In order to re-prioritise peace and avoid ASI-directed or enabled conflict, we must prove that we want a peaceful world. Peaceful actions include conventions, treaties and agreements, with a Universal Global Peace Treaty being essential to revitalize the United Nations. This concrete action would communicate to a future ASI that our priority is peace, enabling partnership and constraining the risk of war.
    As part of this research, we drafted a Cyberweapons and Artificial Intelligence Convention, which we hope can be used as a prelude to the ASI component of the Universal Global Peace Treaty. This is a first step to build on.
    A Universal Global Peace Treaty could be brought to the bargaining table when the Russia-Ukrainine war is over in an attempt to revitalize the UN system.
    We need more regulation on the development of artificial superintelligence. This would help to communicate to politicians that their development is safe and get them on board whilst dissuading them from subverting AGI. Likewise, it would also ensure that development is safe by advancing a peace-based AI infrastructure, one tied to an international treaty-based approach, the UGPT. We need principles-based approaches like Eliezer Yudkowsky’s concept of Friendly AI to also embrace international relations theory., especially concepts that might enable 'best practices' for human-ASI relations, like conforming instrumentalism.
    In historical terms, there exists only a very short window of time for countries to sign a treaty. A draft of the treaty does already exist, but this needs to be progressed urgently, certainly before 2045, many companies' goal for achieving AGI. One idea to accelerate this process is to find a mechanism to enable every country to understand that, as with the race to draft and sign the Geneva Conventions, it is in a race to shape the UGPT and so the principles that the ASI might be persuaded to adopt.

      Let your research make a social impact

      About this research

      This Journal Article was part of a collaborative effort

      Elias Carayannis

      This research was independently conducted and did not receive funding from outside of the university.

      Recommended for

      About this research

      This research was independently conducted and did not receive funding from outside of the university.

      This paper was co-authored

      Elias Carayannis

      Recommended for

      Background

      With the increasing application of artificial intelligence (AI) to warfare, attempts to develop advanced applications of artificial superintelligence (ASI) in this field are almost inevitable. At the moment, ASI precursor technology, Artificial General Intelligence (AGI), is, for the most part, not being engineered for warfare; instead, it is being created for civilian purposes. However, because of its potential for technological supremacy, once achieved ASI will certainly either be applied to warfare or developed for warfare.

      AGI technology is being developed by many projects globally, but Silicon Valley and China are the closest. It is certain that AGI will be informed by principles and values. For instance, if Russia were to be the first to develop this technology, then it could be informed by Cosmism (or Putinism). Given Russian expansionist nationalism involves ‘hot’ war, a Cosmist ASI could direct a global war.

      Preloading principles into an ASI thus seems essential. For instance, if an emerging ASI endorses a ‘Star Trek module’, it could adopt an optimistic vision of how to develop Earth, ourselves, the solar system, and the galaxy, where one or more ASIswork together with humanity as friendly partners.

      As such, what Eliezer Yudkowsky and other Silicon Valleyers are trying to do is to hard code being ‘friendly’ into an ASI such that it could never turn on humanity, i.e., by hard coding it to always act altruistically, via high-level philosophical principles, or ‘supergoals’.

      The obvious threat is that an ASI could still be used to wage catastrophic war by one party on another through political subversion of its supergoals. Turchin and Denkenberger’s position is that politico-military subversion of an ASI will always be attempted; this is also our position in this paper as nothing at present prevents such subversion. With this context in mind, a ‘Terminator’ situation is a possible worst case scenario. Hot, and even possibly cold wars, could indicate to an ASI that countries want war and violence over peace, and ‘winning’ a game of global war could be what the ASI understands to be its own role.

      With regard to peacebuilding, the current UN system has not been good enough at preventing major conflicts, especially if we bear in mind the latest Russia-Ukraine war and the possibility of larger scale warfare arising from the New Cold War.

      The end goal of peacebuilding with regard to international relations and humanity’s relations with an ASI is a universal, global peace. There already exists a draft of a Universal Global Peace Treaty, drafted by the Center for Global Nonkilling, which would be a diplomatic symbol for peace understandable by an ASI, preferably if its creating nation state were a signatory, and especially if it were also a signatory, as the rules and principles of diplomacy would hopefully apply. The closest we have come so far to such an instrument is the Global Ceasefire in 2020, which came from seemingly nowhere, due to COVID, and was backed by the UN.

      What diplomatic rules could be deployed are also discussed in this article. We suggest conforming instrumentalism, the underpinning dynamic for the postwar era Geneva Conventions and likely for the UN itself.

      What findings means

      One aim of a Universal Global Peace Treaty (UGPT) would be to demonstrate, even if symbolically, that peace is a top-level value or ‘supergoal’ for humanity so that a future artificial superintelligence (ASI) is also hardcoded with, or at least is aware of the importance of, this principle.

      Our position is that hard coding ‘peace’ as a supergoal in humanity via the UGPT would inform an ASI that peace is also its own goal – which would mean humanity could have as an ally an ASI who is making decisions on the basis of wanting peace. This could help to ensure that if an ASI were to learn to act independently of its creator, that is, to become, agential, then it would do so with a peaceful objective.

      Based on conforming instrumentalism theory, an ASI would conform towards peace if this is hardcoded from the very beginning and demonstrated to be a top-level global civilizational ideal through countries signing and ratifying the UGPT.

      However, this will not work if peace is slotted in at the very last moment nor if pursuing peace as an ideal is counter-factual. Before we have an all-powerful ASI that might be able to malignly control, or target, humanity, we want it to truly believe that peace is a priority for us and so that cooperative, even friendly relations are not just possible but preferable.

      Findings in practice

      One aim of a Universal Global Peace Treaty (UGPT) would be to demonstrate, even if symbolically, that peace is a top-level value or ‘supergoal’ for humanity so that a future artificial superintelligence (ASI) is also hardcoded with, or at least is aware of the importance of, this principle.

      Our position is that hard coding ‘peace’ as a supergoal in humanity via the UGPT would inform an ASI that peace is also its own goal – which would mean humanity could have as an ally an ASI who is making decisions on the basis of wanting peace. This could help to ensure that if an ASI were to learn to act independently of its creator, that is, to become, agential, then it would do so with a peaceful objective.

      Based on conforming instrumentalism theory, an ASI would conform towards peace if this is hardcoded from the very beginning and demonstrated to be a top-level global civilizational ideal through countries signing and ratifying the UGPT.

      However, this will not work if peace is slotted in at the very last moment nor if pursuing peace as an ideal is counter-factual. Before we have an all-powerful ASI that might be able to malignly control, or target, humanity, we want it to truly believe that peace is a priority for us and so that cooperative, even friendly relations are not just possible but preferable.

      Methodology

      The research was guided by an international relations analytical lens called conforming instrumentalism, and the aim was to try to understand whether there could be a positive influence on global peacebuilding through a Universal Global Peace Treaty (UGPT). It was based on desk-based research, a literature review and the application of the analytical lens across the literature. As part of the research deliverables, we also drafted a Cyberweapon and Artificial Intelligence Convention.

      A weakness of this research is that it is that we do not yet have a state sponsor of the UGPT, and some key organisations developing artificial superintelligence are very private about their operations and may not wish to join an advocacy coalition.

      Glossary

      Conforming Instrumentalism
      A new international relations theory that can be used to contextualise possible AGI behaviour that might conform to the world around it by processing its surroundings and instrumentalising the ideas it sees and then conforming, upwards or downwards, to civilizational ideals, preferably the social norms that peacebulding and cooperative humans idealise.

      Share these insights

      Want to read the full paper? It is available open access

      Carayannis, E.G., Draper, J. Optimising peace through a Universal Global Peace Treaty to constrain the risk of war from a militarised artificial superintelligence. AI & Soc (2022).