With Great Power Comes State Responsibility
In the coming years, several Great Powers are expected to gain possession of artificial intelligence which surpasses human individuals in speed, wit, and knowledge. This development is difficult to imagine, almost unthinkable. Yet if playing along with the assumption, what seems even harder to imagine is the possibility that such states would do anything other than leveraging the technology to advance their national interests, or die trying. In this headspace, internationally wrongful acts would be unfortunate, inevitable necessities.
Yet the fact that internationally wrongful acts form a coherent notion is remarkable, and is indicative of the civilizational achievement of developing an international legal order. What makes international law notable is that states, its primary subjects, are recognized as sovereign, ultimate custodians of their territories and citizenry. How can a legal order function if the will of their subjects reigns supreme? The decentralized legal order relies on subjects which consent to be bound by certain rules. Through the voluntary acceptance of legal obligations, coupled with credible redress mechanisms backed by other members of the international community, international law becomes substantive even in the absence of supranational authorities per se.
Two categories of precedents are worth mentioning. First, states are routinely accepting the jurisdiction of international courts on various matters. This may happen through compromissory clauses in treaties, where parties consent to having disputes adjudicated by a particular international court (e.g., UN Convention on the Laws of the Sea). This may also happen through optional clause declarations in the context of the Statute of the International Court of Justice. Such a conditional unilateral act implies the state’s consent to accepting ICJ’s jurisdiction in advance on a broad range of matters for disputes with states which have made equivalent declarations (e.g., UK, Japan, India, etc.). Perhaps unparalleled in scope among intergovernmental organizations are the binding resolutions of the UN Security Council, which when it deems peace to be threatened, can for instance mandate members to impose sanctions on the destabilizing party. It would not be difficult to imagine a world less willing to practice the pacific resolution of disputes by deferring to institutions designed with an aspiration of impartiality and sovereign equality.
Second, besides sovereign delegation of jurisdiction, another notable precedent consists in the designation of jus cogens. The formal sources of international law include, among others, established custom (i.e., regularities in historical international practice) and general principles (i.e., regularities across legal systems). Yet derogation from these is often possible through conventions which explicitly overturn particular norms for their parties. That said, a select few norms, typically derived from customary law and general principles, have ascended to jus cogens status, meaning that derogation from them is prohibited. These include the prohibition of torture and the right to self-determination. In effect, jus cogens norms collectively reflect the recognition of a shared normative body, one deemed to be timeless and fundamental. It would not be difficult to imagine a world of sovereigns less willing to cede legitimacy of normative foundations to the broader community.
Of course, the international legal order is far from perfect. Its law often gets breached, decisions and processes are often plagued by double standards, and weaker subjects have limited countermeasures when injured. As a triple example, consider the lack of substantive response from the UNSC to Russian aggression in Ukraine, partly due to Russia having continued the legal personality of USSR, and thus maintained a permanent seat which grants it veto rights on resolutions. Yet writing off international law as wholly vacuous would also be wrong, given the thousands of international disputes which have been formally settled through adjudication or arbitration, and which may have counterfactually been settled more violently. That said, one could also argue that a decentralized legal order is only workable in a “balance of power” setting, with few remedies against the internationally wrongful acts of a hegemon with a decisive strategic advantage, even when breached norms are erga omnes. Law legitimizes certain uses of power and delegitimizes others, and so has an intrinsic political dimension that would be rewritten by victors.
Speaking of decisive strategic advantage, there is a case to be made that much technical work on avoiding seizure of power by artificial intelligence is inadvertently contributing to seizure of power by its creators. Let outer alignment refer to the process of imbuing such virtual entities with a particular idiosyncratic worldview, and let control refer to the process of extracting value from such virtual populations even if their worldview is not aligned with one’s own. Indeed, by training them in complacency or by constraining their behavior, it seems less likely that they would succeed in executing a coup d’etat. Yet the same techniques could be used to weaponize these inherently-complacent or tightly-controlled societies of virtual entities, and exercise aggression over all other sovereigns. Hard law predicated on obligations of result already exist to delegitimize aggression against sovereigns, regardless of the novelty of the means, but the question is whether such law would remain substantive under such power dynamics.
Before dreaming of alternatives to the false dichotomy, it is worth discussing what is the point. We are about to instantiate beings capable of having undue influence on the world. What are we after? Is it the supremacy of one people? That feels extremely arbitrary and unfounded. Is it the persistence of one species? That also feels quite arbitrary. Is it the flourishing of all sentient beings? That feels less arbitrary, but have we given it enough thought? One might argue that the most defensible thing to do is, on the one hand, to bring into the world minds which arduously reflect on what is the most defensible thing to do, and on the other hand, to coordinate against illegitimate use of power which might irremediably prevent that from happening. While outer alignment might help point at the notion of defensibility, it may be the case that it can be directly operationalized in closed-form, and potentially without the use of control techniques.
How would an international arrangement inspired by these ideals look like? The part concerned with preemptive remedies could be pursued through somewhat familiar means, namely prohibition of practices conducive to threats to peace. For instance, the consumption of dual-use capabilities, such as autonomous hacking or autonomous research, could be metered and managed through arms control doctrines. When it comes to the proliferation of defensibility, capabilities concerned with reasoning about lawful behavior could be subsidized or positively mandated. In a sense, the ICJ’s 15 judges of different nationalities, representing the “main forms of civilization […] of the world,” represent an attempt at machinery capable of impartially determining the most defensible position among parties in a dispute. The facts of the case, the growing body of international law, together with the interpretative abilities of the court, make the adjudication possible. As with most forms of cognitive labor, such processes may also become automated and commoditized, yet scaling up moral reasoning in particular may help infuse the actions of machines with justice.
Questions about how we ought to handle imminent virtual populations are delicate and important, and we have barely scratched the surface. Technical work should stem from an intentional approach to these questions, or at least be designed to productively accommodate a broad range of stances. Addressing such questions will likely require drawing from a spectrum punctuated by global governance, international relations, diplomacy, international law, philosophy of law, moral epistemology, and metaethics. For further discussion on related ideas along this spectrum, refer to current work on virtual diplomacy and previous work on computational philosophy.