AI for Peace

AI Peace

This article was submitted in response to the call for ideas issued by the co-chairs of the National Security Commission on Artificial Intelligence, Eric Schmidt and Robert Work. It addresses the fourth question (part a.) which asks what international norms for artificial intelligence should the United States lead in developing, and whether it is possible to create mechanisms for the development and enforcement of AI norms.


In 1953, President Dwight Eisenhower asked the world to join him in building a framework for “Atoms for Peace.” He made the case for a global agreement to prevent the spread of nuclear weapons while also sharing the peaceful uses of nuclear technology for power, agriculture, and medicine. No one would argue the program completely prevented the spread of weapons technology: India and Pakistan used technology gained through Atoms for Peace in their nascent nuclear weapons programs. But it made for a safer world by paving the way for a system of inspections and controls on nuclear facilities, including the establishment of the International Atomic Energy Agency and, later, the widespread ratification of the Treaty on the Nonproliferation of Nuclear Weapons (NPT). These steps were crucial for building what became known as the nuclear nonproliferation regime.

The world stands at a similar juncture today at the dawn of the age of artificial intelligence (AI). The United States should apply lessons from the 70-year history of governing nuclear technology by building a framework for governing AI military technology.



What would “AI for Peace” look like? The nature of AI is different than nuclear technology, but some of the principles that underpinned the nonproliferation regime can be applied to combat the dangers of AI. Government, the private sector, and academia can work together to bridge national divides. Scientists and technologists — not just traditional policymakers — will be instrumental in providing guidance about how to govern new technology. At a diplomatic level, sharing the peaceful benefits of technology can encourage countries to open themselves up to inspection and controls. And even countries that are competitors can cooperate to establish norms to prevent the spread of technology that would be destabilizing.

AI for Peace could go beyond current efforts by involving the private sector from the get-go and identifying the specific dangers AI presents and the global norms that could prevent those dangers (e.g., what does meaningful human control over smart machines mean in specific contexts?). It would also go beyond Department of Defense initiatives to build norms by encompassing peaceful applications. Finally, it would advance the United States’ historic role as a leader in forging global consensus.

The Dangers of Artificial Intelligence

The uncertainty surrounding AI’s long-term possibilities makes it difficult to regulate, but the potential for chaos is more tangible. It could be used to inflict catastrophic kinetic, military, and political damage. AI-assisted weapons are essentially very smart machines that can find hidden targets more quickly and attack them with greater precision than conventional computer-guided weapons.

As AI becomes incorporated into society’s increasingly autonomous information backbone, it could also pose a risk of catastrophic accidents. If AI becomes pervasive, banking, power generation, and hospitals will be even more vulnerable to cyberattack. Some speculate than an AI “superintelligence” could develop a strategic calculating ability so superior that it destabilizes arms control efforts.

There are limits to the nuclear governance analogy. Whereas nuclear technology was once the purview only of the most powerful states, the private sector leads AI innovation. States could once agree to safeguard nuclear secrets, but AI is already everywhere — including in every smartphone on the planet.

Its ubiquity shows its appeal, but the same ubiquity lowers the cost of sowing disorder. A recent study found that for less than $10 anyone could create a fake United Nations speech credible enough to be shared on the internet as real. Controlling the most dangerous uses of technology will require private sector initiatives to build safety into AI systems.

Scientists Speak Out

In 2015, Stephen Hawking, Peter Norvig, and others signed an open letter calling for more research on AI’s impacts on society. The letter recognized the tremendous benefits AI could bring for human health and happiness, but also warned of unpredictable dangers. The key issue is that humans should remain in control. More than 700 AI and robotics researchers signed the 2017 Asilomar AI Principles calling for shared responsibility and warning against an AI arms race.

The path to governing nuclear technology followed a similar pattern of exchange between scientists and policymakers. Around 1943, Niels Bohr, a famous Danish physicist, made the case that since scientists created nuclear weapons, they should take responsibility for efforts to control the technology. Two years later, after the first use of nuclear weapons, the United States created a committee to deliberate about whether the weapons should become central to U.S. military strategy, or whether the country should forego them and avoid a costly arms race. The Acheson-Lilienthal committee’s proposal to put nuclear weapons under shared international control failed to gain support, but it was one step in a consensus-building process. The U.S. Department of Defense, Department of State, and other agencies developed their own perspectives, and U.N. negotiations eventually produced the Treaty on the Non-Proliferation of Nuclear Weapons (NPT). Since entering into force in 1970, it has become the most widely subscribed arms control treaty in history with a total of 191 signatory states.

We are in the Acheson-Lilienthal age of governing AI. Neither disarmament nor shared control is feasible in the short term, and the best hope is to limit risk. The NPT was created with the principles of non-possession and non-transfer of nuclear weapons material and technology in mind, but AI code is too diffuse and too widely available for those principles to be the lodestar of AI governance.

What Norms Do We Want?

What then does nonproliferation look like in AI? What could or should be prohibited? One popular proposal is a no kill rule for unassisted AI: humans should bear responsibility for military attack.

A current Defense Department directive requires “appropriate levels of human judgment” in autonomous system attacks aimed at humans. This allows the United States to claim the moral high ground. The next step is to add specificity to what appropriate levels of judgement means in particular classes of technology. For example, greater human control might be proportional to greater potential for lethality. Many of AI’s dangers stem from the possibility that it might act through code too complex for humans to understand, or that it might learn so rapidly so as to be outside of human direction and therefore threaten humanity. We must consider how these situations might arise and what could be done to preserve human control. Roboticists say that such existing tools as reinforcement learning and utility functions will not solve the control problem.

An AI system might need to be turned off for maintenance or, crucially, in cases where the AI system poses a threat. Robots often have a red shutdown button in case of emergency, but an AI system might be able learn to turn off its own off switch, which would likely be software rather than a big red button. Google is developing an off switch it terms a “kill switch” for its applications, and European lawmakers are debating whether and how to make a kill switch mandatory. This may require a different kind of algorithm than currently exists — one with safety and interpretability at the core. It is not clear what an off switch means in military terms, but American-Soviet arms control faced a similar problem. Yet arms control proceeded though technical negotiations that established complex yet robust command and control systems.

Building International Consensus

The NPT was preceded by a quarter century of deliberation and consensus building. We are at the beginning of that timeline for AI. The purpose of treaties and consensus building is to limit the risks of dangerous technology by convincing countries that restraint is in the interests of mankind and their own security.

Nuclear nonproliferation agreements succeeded because the United States and the Soviet Union convinced non-nuclear nations that limiting the spread of nuclear weapons was in their interest — even if it meant renouncing weapons while other countries still had them. In 1963, John F. Kennedy asked “…what it would mean to have nuclear weapons in so many hands, in the hands of countries large and small, stable and unstable, responsible and irresponsible, scattered throughout the world.” The answer was that more weapons in the hands of more countries would increase the chance of accident, proxy wars, weak command and control systems, and first strikes. The threat of nuclear weapons in the hands of regional rivals could be more destabilizing than in the hands of the superpowers. We do not yet know if the same is true for AI, but we should investigate the possibility.

Access to Peaceful Technology

It is a tall order to ask countries to buy into a regime that limits their development of a powerful new technology. Nuclear negotiations offered the carrot of eventual disarmament, but what disarmament means in the AI context is not clear. However, the principle that adopting restrictions on AI weapons should be linked to access to the benefits of AI for peaceful uses and security cooperation could apply. Arms control negotiator William Foster wrote in 1967 that the NPT treaty would “stimulate widespread, peaceful development of nuclear energy.” Why not promise to share peaceful and humanitarian applications of AI — for agriculture and medicine, for example — with countries that agree to participate in global controls?

The foundation of providing access to peaceful nuclear technology in exchange for monitoring materials and technology led to the development of a system of inspections known as safeguards. These were controversial and initially not strong enough to prevent the spread of nuclear weapons, but they took hold over time. A regime for AI inspection and verification will take time to emerge.

As in the nuclear sphere, the first step is to build consensus and identify what other nations want and where common interest lies. AI exists in lines of code, not molecules of uranium. For publicly available AI code, principles of transparency may help mutual inspection. For code that is protected, more indirect measures of monitoring and verification may be devised.

Finally, nuclear arms control and nonproliferation succeeded as part of a larger strategy (including extended deterrence) that provided strategic stability and reassurance to U.S. allies. America and the Soviet Union — despite their Cold War competition — found common interests in preventing the spread of nuclear weapons. AI strategy goes hand-in-hand with a larger defense strategy.

A New AI for Defense Framework

Once again, the world needs U.S. global leadership — this time to prevent an AI arms race, accident, or catastrophic attack. U.N.-led discussions are valuable but overly broad, and the technology has too many military applications for industry alone to lead regulation. Current U.N. talks are preoccupied with discussion of a ban on lethal autonomous weapons. These are sometimes termed “killer robots” because they are smart machines that can move in the world and make decisions without human control. They cause concern if human beings are not involved in the decision to kill. The speed and scale of AI deployment calls for more nuance than the current U.N. talks can provide, and more involvement by more stakeholders, including national level governments and industry.

As at the dawn of the nuclear age, the United States can build global consensus in the age of AI to reduce risks and make the world safe for one of its leading technologies — one that’s valuable to U.S. industry and to humanity.

Washington should build a framework for a global consensus on how to govern AI technology that could be weaponized. Private sector participation would be crucial to address governance, as well as how to share peaceful benefits to incentivize participation. The Pentagon, in partnership with private sector technology firms, is a natural leader because of its budget and role in the industrial base.

An AI for Peace program should articulate the dangers of this new technology, principles (e.g. no kill, human control, off switch) to manage the dangers, and a structure to shape the incentives for other states (perhaps a system of monitoring and inspection). Our age is not friendly to new treaties, but we can foster new norms. We can learn from the nuclear age that countries will agree to limit dangerous technology with the promise of peaceful benefits for all.



Patrick S. Roberts is a political scientist at the nonprofit, nonpartisan RAND Corporation. Roberts served as an advisor in the State Department’s Bureau of International Security and Nonproliferation, where he worked on the NPT and other nuclear issues.

Image: Nuclear Regulatory Commission