Earlier this month, US researchers used AI to design new viruses for the first time. Though these viruses were created only to combat bacteria like E. coli, the breakthrough has led to increased calls for international agreements to regulate the use of AI to develop biological weapons.
However, one drawback of our current unipolar moment is that the US has become unaccustomed to negotiating formal agreements with its opponents. In any case, there is currently a lack of interest in formal agreements.
What would it take to actually get agreement on AI?
First, there must be a shared perception of risk and threat. The Cuban Missile Crisis inspired a generation of agreements for nuclear weapons, conventional military action, and WMD. Leaders in both Russia and the United States remembered their experiences in World War II and feared the return of global war fueled by nuclear weapons. This common understanding created the basis for negotiation and, ultimately, agreement.
Second, there must be agreement among all parties that this risk needs to be addressed. China and Russia appear unwilling to engage with the US in serious discussion on AI’s potential risks. Russia’s thinking is unclear, and China, while it has expressed concern about the risk of AI’s weaponization, seems to believe it can manage risk without pausing development. Unilateral export controls are not a replacement, and technical guardrails still require political consent from all parties. Legislation can help only if it adds credibility to US proposals.
Third, there must be recognition of the limits of participation. Frankly, more participants make meaningful agreement less likely. The United Nations is not the place to initiate discussions, given its diffuse membership. The temptation will be to replicate with China the bilateral format of the Cold War, but Russia, Europe, and perhaps India will want to be involved. With every additional participant, the difficulty of obtaining agreement grows. We should also avoid a Davos-style format with well-meaning private sector leaders participating as equals with states. Non-governmental participation risks turning negotiations into seminars.
International agreements on AI are not possible without buy-in from the national leadership of the major nuclear powers. Any agreement on AI that does not include China is of limited value. However, the Chinese may be tempted to enter into negotiations if they believe it can be used to hamper US progress. There is precedent for this in both Soviet and Chinese behavior. China, for example, was always careful in conventional arms talks to propose thresholds that covered US weapons but not Chinese ones.
Controls on biological weapons are difficult to design without being unnecessarily restrictive because of their dual-use applications. Efforts to restrict biological weapons and their manufacture go back to the 1960s. Earlier efforts even considered restrictions on breweries, since the equipment used for brewing can also be used to produce biological weapons, and remote verification is difficult without consent. In one extreme example, the US inadvertently bombed a pharmaceutical plant in Sudan after satellite imagery and samples of nearby dirt seemed to provide evidence that it was involved in the production of biological weapons.
Biological weapons are easy to create, but hard to use. While stockpiles have been assembled (particularly by Russia), no one has used a biological weapon (other than in a salad bar and a subway car). They’re easy to make, but unpredictable and difficult to deploy; they’re just as likely to turn around and infect the attacker. AI makes it easier to design more sophisticated weapons, but it does not change the disincentives for use or the operational difficulties that use entails.
Managing AI risk requires asking what is needed to supplement existing agreements, such as the 1972 Biological Weapons Convention (BWC), which already restricts the development of biological weapons. A successful agreement could take two forms.
- A general commitment at the political level not to engage in research that uses AI to develop advanced biological weapons. This requires buy-in at the presidential level from China, Russia, and the United States, which appears unlikely at this time. While arms control was always a summit agenda item in discussions between the Soviet Union and the US after 1960, and security was always on the agenda for summits between the US and China since the first discussions between Mao and Nixon, China and Russia are preoccupied, and this Administration prefers to focus on trade and business.
- The second is to develop and agree on specific performance thresholds. For example, the Missile Technology Control Regime uses both payload and range specifications. Our understanding of AI is not yet sufficient to allow us to develop durable thresholds, complicating the task of reaching agreement. Blanket bans on technology are so appealing to commentators and Congress because they do not require the hard work of identifying the actual specifics of risk.
At a minimum, this suggests that calls for international agreement are premature, an expression of emotion and concern rather than an effective policy proposal. AI is not a weapon. It can improve existing weapons and maybe create new ones, but that is currently best addressed by amending existing non-proliferation agreements. We may never need an agreement specifically for AI. Add to this the difficulties of verification, and we are far from the point where international AI agreements could improve international security.
James Lewis is a Distinguished Fellow on the Tech Policy Program at the Center for European Policy Analysis.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Tech 2030
A Roadmap for Europe-US Tech Cooperation
