This is an excerpt from the Foresight 2019 AGI Cooperation Report. At the workshop, I led a session on the impact of AI on nuclear escalation paths. Here I summarize the findings from the session.
Advances in military-relevant AI technologies threaten to worsen risks of accidental nuclear escalation. Perhaps the greatest risk comes from the uncertain effectiveness of these capabilities, both in theory and in application. Despite their power, these technologies should not be seen as a “game changer,” but should be understood in the context of existing and in-development weapons and deployment capabilities. While we believe the most likely effect of these technologies will be to increase the risk of inadvertent escalation, there is an opportunity to use these technologies to accomplish the opposite, and reduce these risks.
Since the invention of nuclear weapons, competition for nuclear advantage has driven advances in three areas. The first is the ability to reliably, accurately, and quickly deliver nuclear weapons to their targets. The second is the ability to effectively monitor nuclear developments in other countries, including the development and tests of nuclear weapons, the deployment of nuclear weapons, and, crucially, the launch or use of nuclear weapons. The third category, the ability to directly defend one’s The Need for Cooperation in AI Geopolitics 9 Artificial General Intelligence: Toward Cooperation weapons, and, crucially, the launch or use of nuclear weapons. The third category, the ability to directly defend one’s country from nuclear attack, has also advanced, though carries far less relevance to a great power conflict due to the fundamental difficulty of missile defense. AI technology will impact all three areas.
The greatest risks to strategic stability come from uncertainties about all three categories of development, with the greatest risks stemming from the first, the ability to reliably, accurately, and quickly deliver nuclear weapons to their targets. On the other hand, concrete advances in the second category, the ability to effectively monitor nuclear developments of states around the world and detect weapons tests and missile launches, has the potential to improve strategic stability. Improving the reliability of launch detection systems would reduce the risk of accidental escalation, and improving the detection of weapons development could improve confidence in nuclear arms control agreements.
Benefits & Advantages
The greatest potential advantage for AI technology lies in improving detection capabilities – that is, the ability to effectively monitor nuclear developments in other countries, including the development and tests of nuclear weapons, the deployment of nuclear weapons, and, crucially, the launch or use of nuclear weapons. More reliable missile launch detection capabilities could reduce the risk of accidents caused by the current launch detection systems which have previously exhibited near-disastrous false positives. Greater monitoring capability of nonnuclear nations can help prevent further nuclear proliferation and lead to better enforcement of the Treaty on the Non-Proliferation of nuclear weapons (NPT). Improved ability to monitor the nuclear developments of current nuclear states may allow for greater confidence in existing and future arms control agreements.
Risks & Obstacles
The greatest risks to strategic stability and the prevention of nuclear escalation come from two main factors. The first is uncertainty of nuclear defensive and offensive capabilities, which can worsen arms races and increase pressure to act in the fog of war. The resulting missile build up in the US due to the perceived “missile gap” during the Cold War demonstrates how inadequate intelligence about another state’s nuclear capabilities can worsen arms race dynamics. If deployment of existing AI technologies lead to improved intelligence, reconnaissance, and surveillance (ISR), the net effect should be to reduce the states’ uncertainties regarding each other’s nuclear capabilities.
The second factor is the ability for AI technologies to increase the speed and stealth of weapons by enabling enhanced autonomous capabilities. We expect that advances in autonomous vehicles, especially drones, will lead to systems with greater reaction times and an increased capacity for rapid conflict escalation. This would decrease the time for military leadership to react in a crisis and increase the incentive for preemptive escalation. Furthermore, a natural response to the threat of enemy autonomous weapons systems will be an arms race in these types of systems.
The application of military-relevant AI technologies is full of uncertainty. We cannot foresee exactly how the technologies will develop nor understand their full impact. The same military uncertainties that heighten risk of accidental escalation also make the development of public policy on these topics quite challenging. We believe the only way to make progress in this domain is to improve the quality of research and discussion and the pipeline that connects this discussion to actual policy. The Need for Cooperation in AI Geopolitics 10 Artificial General Intelligence: Toward Cooperation When nuclear weapon systems were first developed in the 1940s and 1950s in the United States, public and private collaboration was much stronger than it is today. Unlike that era, most advances in AI technologies now come from private companies, often in collaboration with academic institutions. As a result, policy makers in the US government and military suffer from a lack of expertise in the technology their policies target. We recommend that researchers and private institutions with active safety research programs and a strong commitment to the safe application of AI technologies establish relationships with policy makers to improve mutual understanding of technology and policy challenges.
Ultimately, reducing risk from nuclear escalation has a large impact on all countries in the world. While all militaries have incentives to hide and protect strategic secrets, there are also reasons to collaborate on understanding the strategic ramifications of new technologies. No country in the world would benefit from nuclear escalation. We hope that states will choose to develop AI technologies in areas where they will improve strategic stability while working towards international agreements to limit destabilizing applications of these technologies.
We recommend a thorough study of existing and near-horizon autonomous weapons systems and their impact on nuclear escalation paths. So far, this section has avoided discussion of AGI due to the speculative nature of AGI systems in contrast to the comparatively better specified nature of current AI technologies and their application to nuclear defense and offensive systems. The one area in which AGI appears relevant is how perceptions of AGI development may affect nuclear stability and the deterrence of nuclear first strikes. Even if AGI capabilities are far out, the perception of a state’s nearness to decisive strategic advantage via AGI capabilities could worsen strategic stability. We recommend cautious investigation of the impact of perceived AGI capabilities on strategic stability