This is the second piece of a blog post series that explores how AI prediction services affect the risks of war. It is based on my 10-week summer research project at Stanford Existential Risk Institute. See my first post for a summary of my project and my epistemic status.
In the first post, I have surveyed different prediction technologies, how they might be relevant to nation-state governments, and what their development trajectories might look like. This post applies international relations literature to examine what implications follow for the risk of war. The next post will describe three possible “world order” scenarios after prediction capability takes off.
Risks of War
I choose to focus on the rationalist explanations of war within international relations literature. They are not perfect explanations, but they seem to me plausible enough to be a starting point.
If we assume that most wars are wasteful, and there is usually a range of peaceful ways to resolve the conflicts that both sides would prefer than going to war, why do people still go to war?
One explanation is that, when pondering whether to go to war, leaders have private information about their military capabilities and the costs of war. They thus have an incentive to bluff for better deals. This makes it harder for them to reach peaceful agreements. Costly wars can be means by which nation-states screen each other’s military capability and reservation prices. Empirically, nation-states with reconnaissance satellites, which reduce private information about military capabilities and the exploitability of surprise attacks, are significantly less likely to be involved in high-casualty militarized disputes (MIDs). This seems to weakly support the “private information” thesis, suggesting a possible pacifying effect of advanced prediction capability.
However, I have come to be skeptical of this mechanism. Nation-states still have mixed incentives: they want to avoid costly wars while striking better deals. Therefore, they have an incentive to create uncertainty and defy the other’s prediction capability. It seems like military capabilities and costs of war are hard to predict but easy to misrepresent. Therefore, an increase in prediction capability does not necessarily lead to less private information - just as nation-states may seek new forms of secrecy and espionage, like hacking, after satellites make the physical world more transparent.
We could roughly distinguish wars into two types: (1) great power vs. small power (e.g. governments vs. insurgents, the US vs. Iran); (2) great power vs. great power (e.g. the US vs. China). We assume that great power benefits from a large boost in militarily relevant prediction capability and small power does not. It seems that, speaking only of the “private information” mechanism and holding other things constant, the probability of war would decrease for the first dyad but not change much for the second.
Cost of War
Wars are not always wasteful. Nor are they invariably wasteful. At a systemic level, we can think of the general probability of war as an equilibrium result of mixed strategies chosen by nation-states. They first consider the expected utility of wars relative to other ways they could deal with each other. Then they deliberately choose a positive probability of war, perhaps by arming at a certain level, such that they are indifferent between war and other peaceful ways of doing international politics. If so, AI prediction services could change the probability of war by changing the structure of the payoffs.
The cost of war would be driven by many political, economic, and technological forces other than prediction capability advanced by AI. But it may still be worth it to discuss some ways in which this could affect the cost of war, because:
- It helps decompose the big, abstract problem;
- It helps understand what kinds of prediction services we want to push forward or refrain from development.
I tend to believe that AI prediction services will confer a defensive advantage in the next 5 to 15 years.
- Easy defense. Current applications such as early warning systems and defense planning using computational game theories seem to decrease the relative cost of deterrence or defense and, when holding the investment ratio in defense and offense constant, lower the probability of a successful surprise attack, at least in the physical world. That said, my research based on open-source information is likely biased towards finding defensive applications as opposed to offensive ones. This is because nation-states have an incentive to make public the former to deter the enemy while hiding the latter.
- Hard offense. Gaining a decisive offensive advantage by accurately predicting the adversary's military actions in depth is difficult and probably necessitates aggressive data collection from the adversary, while a possible offensive advantage from predicting a real-time vulnerability - “seeing through the fog of war” - may be quickly offset by a randomization strategy adopted by the adversary. However:
- Many warn that AI prediction services could be used to gauge locations of nuclear arsenals, threaten the adversary’s second-strike capability, and thereby favor the offense. This seems particularly relevant given US military intelligence predominance vis-a-vis its foreign nuclear adversaries. My skepticism of such use-case being influential mainly consists in (1) the seeming difficulty to verify the accuracy of prediction without risking nuclear exchanges; (2) the seeming easiness of adversaries to create new nuclear weapons to gain its edge.
- Small powers can be less capable of adaptive randomization. If so, AI prediction services can increase the power asymmetry between small power and great power, decreasing the cost of the latter attacking the former.
- Uncertainty about implementation effectiveness. In general, the actual usefulness of incorporating prediction technologies into military operations would closely depend upon organizational capacities, including the ability to collect intelligence from the adversary and to manage complex information systems. New technologies can sometimes compromise military effectiveness due to ineffective management and ill-adaptation.
I think AI prediction services could make international institutions cheaper. If so, wars would become relatively more expensive.
- Cheaper enforcement, evaluation, and generation of contracts. Through various data analysis pipelines, AI prediction services can provide cheap, real-time monitoring, better evaluation of distributional consequences, smart suggestions of side payments, issue linkages, and new agreements to address unanticipated consequences.
- More trade-like agreements. Nation-states often have mixed incentives when cooperating: they have an incentive for reaching an agreement, but they may prefer different types of agreement, or they may prefer to cheat and lie without enforcement and monitoring mechanisms. Thick international regimes, like the UN and the WTO, arise when supply meets demand: international regimes reduce uncertainty and risk by linking discrete issues and developing overarching norms and principles around issue areas; great powers supply the coercive power backing international regimes, capturing private gains minus organization costs. By supplying cheaper enforcement, AI prediction services could reduce the demand for those regimes altogether. There is less need to bundle issues into generic issue areas. Adaptive, issue-specific policy-making is now possible. This new way of doing politics is particularly attractive to nation-states in issue areas where the distribution of gains is sensitive to shocks and uncertainty such as security, environment, and trade. They have short-term agreements, for which they review and renegotiate cheaply, or make contingency contracts.
- Increased, more open membership. When enforcement is costly and uncertainty about each party’s willingness to participate is high, international agreement tends to have more restrictive membership. AI prediction services could reduce such uncertainty. Seemingly suspicious, risky participants need not be excluded but provided with unfavorable terms along with an opportunity of revising terms based on near-term performance. Fastened feedback loop makes misrepresentation extremely costly and therefore mostly unworthy. Public policy coordination on metropolitan scales may also become more common. Such coordination could function as quick, cheap tests for evaluating policy options on larger scales. In consequence,
- The value of comprehensive, long-term partnerships and alliances as costly signals for willingness to participate is reduced.
- But exclusive membership could remain important if data-sharing raises security concerns or if data-sharing has a high fixed cost.
- Ambiguous effect on the level of centralized control international institutions have over nation-states. On the one hand, reduced uncertainty about compliance reduces the need for centralization; reduced uncertainty about distributional consequences raises domestic audience costs of centralization and likely creates a demand for domestic control in democracies. On the other hand, the cost of centralization declines with increased membership and reduced transaction cost; the amount of hegemonic rent that great power as “unitary actors” can extract also declines, making hegemonic competition less attractive and a centralized mechanism potentially more effective. I suspect that, for policy issues that are operational, mundane, with more converging interests among nation-states, without explicit distributional consequences for the domestic audience, centralization is more likely; for policy issues that are new, non-divisible, with high potential for hegemonic rents, and with clearly diverging distributional consequences for the domestic audience, decentralization is more likely.
Some skepticism that these cooperative benefits may not be realized, or take a long time, or of a small size:
- Clumsy institutions and disincentivized individuals. The above reasoning assumes that international institutions are designed rationally, i.e. actors adjust their interaction with each other in response to cost factors. However, in reality, institutions are often clumsy and less responsive to costs. Also, there seems to be little personal incentive for international bureaucrats or politicians from nation-states to revise the currently ill-functioning international institutions or create new ones. Those benefits might thus remain low-hanging fruits for a long time after technology maturity.
- Politicalized prediction. The above reasoning also assumes that nation-states agree to trust particular AI prediction services. However, they could disagree, run distinct “national” services, and hold on to conflicting evidence about policy consequences - just as the models used to inform pandemic policy became politicalized. The more black-box the prediction, the easier the politicization.
I think AI prediction services could increase the relevance of the domestic audience for foreign policymaking. If so, it complicates the calculation of the cost of war: war can be costly for some yet beneficial for others. How it would tilt the calculation depends.
- Ambiguous effect on citizen-leader information gap. Citizens in democratic societies are rationally ignorant. With regard to war and international crises, they typically have a big informational disadvantage relative to politicians. They use simple heuristics, signaling loyalty to local communities or their nation. Politicians exploit this: sometimes they go to war for the rally-round-the-flag effect; sometimes they take excessive risks from a median voter’s viewpoint to gamble a resurrection for their personal careers.  AI prediction services may shorten the process through which the public find out truths. Irrational, inconsistent beliefs have shorter average lifespans. Politicians are held more accountable. However, AI prediction services may also make it easier for politicians to manipulate public sentiments to advance personal agendas. The exact effect of AI prediction services on the information gap between politicians and citizens may depend on data ownership and regulation as well as the diffusion of prediction services. Increased openness of data, secured data-sharing, and widely diffused prediction services such as personified policy impact analysts are likely to empower citizens relative to politicians.
- Expanded public engagement. Citizens largely do not care about “mundane” foreign policy issues. AI prediction services can make the impact assessment of foreign policy cheap and accessible to each citizen. As a result, more citizens care more about foreign policy; politicians are more responsive to electoral implications of foreign policy and compete over different policy stances. This could mean:
- The return of the silent majority. AI prediction services might increase the influence of large interest groups relative to small ones. For example, many democracies favor protectionism over trade liberation in agricultural sectors, because small interest groups like farmers, who benefit tremendously from protectionist policy, can more effectively mobilize than consumers, each of whom are slightly harmed by it. This collective action problem may be overcome if the overall welfare impacts of trade policy and/or median voter’s preference can be cheaply predicted.
- Changing preferences over policy options. AI prediction services might make nation-states prefer less controversial policy options compared to those with clearly conflicting interests. For example, arming and alliances are both policy options to increase national security. If arming requires higher taxation and has spill-over over the national economy, whereas alliance formation requires a resolution of conflicting interests that is domestically costly, then arming might become more attractive.
- More credible commitment and quicker negotiation. Domestic audiences can be a credible commitment of a national policy stance in the international arena. Political scientists argue that such a mechanism explains democratic peace: under domestic opposition and electorate pressure, democratic leaders only initiate a war that they expect to win and fight harder for war, making democracies unattractive targets for war. AI prediction services reduce uncertainty about domestic electorates’ policy preferences, thereby making commitment credible and international negotiation quicker, for both war or non-war issues. For crisis bargaining, prediction of public preferences may allow the issuing of threats and deterrence to proceed in secrecy, avoiding public confrontation that risks locking in costly escalation and promising face-saving. However, such an effect is conditional on domestic audiences having a resilient consensus over a policy outcome, e.g. peace or military victory. When their preferences are volatile or divided, commitment becomes less credible; shorter contracts are preferred.
Some skepticism that domestic audience will remain irrelevant:
- Still, lack of interest. It may be argued that open-source-based prediction on foreign policy issues has become cheaper, more accurate, and more accessible e.g. via expert consensus, prediction markets, and forecasting platforms. Still, the level of interest from domestic audiences has not correspondingly increased. It could be that previous prediction technologies have not radically decreased the cost of accurate information. If so, AI prediction services may bring the cost down to a threshold and have an observable effect. However, we do not have a good model of when and why domestic audiences care about foreign policy, or policy in general. Their level of interest could be unresponsive to the availability and cost of prediction technologies. If so, AI prediction services may not have much effect at all.
- Again, politicalized prediction. Domestic audiences could fail to agree on which particular AI prediction services to trust, especially for rare events and new situations. If so, prediction could be less distinguishable from persuasion sold by political parties and politicians. Public perception is engineered and does not change much from today.
Bryan R. Early and Erik Gartzke, ‘Spying from Space: Reconnaissance Satellites and Interstate Disputes’, Journal of Conflict Resolution, 23 March 2021, 0022002721995894, https://doi.org/10.1177/0022002721995894. ↩︎
Adam Meirowitz and Anne E. Sartori, ‘Strategic Uncertainty as a Cause of War’, Quarterly Journal of Political Science 3, no. 4 (2008): 327–52. Meirowitz and Sartori’s model shows that nation-states with no ex-ante private information about their military capabilities may choose to create uncertainty, for example by arming with a level of unpredictability and by keeping military secrets about their budgets and particular programs. ↩︎
I drew some cost factors from James D. Fearon, ‘Cooperation, Conflict, and the Costs of Anarchy’, International Organization 72, no. 3 (2018): 523–59, https://doi.org/10.1017/S0020818318000115. ↩︎
Jon R. Lindsay, Information Technology and Military Power, Cornell Studies in Security Affairs (Ithaca ; London: Cornell University Press, 2020). ↩︎
Robert O Keohane, ‘The Demand for International Regimes’, International Organization, n.d., 31. ↩︎
Barbara Koremenos, ‘Contracting around International Uncertainty’, American Political Science Review 99, no. 4 (November 2005): 549–65, https://doi.org/10.1017/S0003055405051877. ↩︎
Barbara Koremenos, Charles Lipson, and Duncan Snidal, ‘The Rational Design of International Institutions’, n.d., 39. ↩︎
Andrea Saltelli et al., ‘Five Ways to Ensure That Models Serve Society: A Manifesto’, Nature 582, no. 7813 (June 2020): 482–84, https://doi.org/10.1038/d41586-020-01812-9. ↩︎
George W. Downs and David M. Rocke, ‘Conflict, Agency, and Gambling for Resurrection: The Principal-Agent Problem Goes to War’, American Journal of Political Science 38, no. 2 (May 1994): 362, https://doi.org/10.2307/2111408. ↩︎
H. E. Goemans and Mark Fey, ‘Risky but Rational: War as an Institutionally Induced Gamble’, The Journal of Politics 71, no. 1 (January 2009): 35–54, https://doi.org/10.1017/S0022381608090038. ↩︎
Matthew A. Baum and Philip B.K. Potter, ‘The Relationships Between Mass Media, Public Opinion, and Foreign Policy: Toward a Theoretical Synthesis’, Annual Review of Political Science 11, no. 1 (June 2008): 39–65, https://doi.org/10.1146/annurev.polisci.11.060406.214132. ↩︎
Christina L. Davis, ‘International Institutions and Issue Linkage: Building Support for Agricultural Trade Liberalization’, American Political Science Review 98, no. 1 (February 2004): 153–69, https://doi.org/10.1017/S0003055404001066. ↩︎
James D. Morrow, ‘Arms versus Allies: Trade-Offs in the Search for Security’, International Organization 47, no. 2 (1993): 207–33, https://doi.org/10.1017/S0020818300027922. ↩︎
See, for example, Bruce Bueno de Mesquita et al., ‘An Institutional Explanation of the Democratic Peace’, American Political Science Review 93, no. 4 (December 1999): 791–807, https://doi.org/10.2307/2586113; Kenneth A. Schultz, ‘Domestic Opposition and Signaling in International Crises’, American Political Science Review 92, no. 4 (December 1998): 829–44, https://doi.org/10.2307/2586306; D. Marc Kilgour, ‘Domestic Political Structure and War Behavior: A Game-Theoretic Approach’, Journal of Conflict Resolution 35, no. 2 (June 1991): 266–84, https://doi.org/10.1177/0022002791035002006. ↩︎
Helen V. Milner and B. Peter Rosendorff, ‘Democratic Politics and International Trade Negotiations: Elections and Divided Government As Constraints on Trade Liberalization’, Journal of Conflict Resolution 41, no. 1 (February 1997): 117–46, https://doi.org/10.1177/0022002797041001006. ↩︎
Very interesting, and I think it mostly goes in the right direction - but I'm not very convinced by the arguments, mostly because I don't think the analysis of causes of war is sufficient here.
For example, even within rational actor models, I don't think you give enough credence to multi-level models of incentives for war, which I discussed a bit here. Leaders often are willing to play at brinksmanship or even go to war because it's advantageous regardless of whether they win. A single case can illustrate: a dictator might go to war to prevent internal dissent, and in that case, even losing the war can be a rallying cry for him to consolidate power. An AI system might even tell people that, but it won't keep him from making the decision if it's beneficial to have a war. And even without a dictator, different constituencies will support or avoid war for reasons unrelated to whether the country is likely to win - because "good for the country overall" isn't any single actor's reason for any decision, and prediction services won't (necessarily) change that.
Thanks for the comment and I enjoy reading the article! I basically agree with what you said and admit that I only get to touch a bit upon this important "multi-level interests problem" within the "domestic audience" section. I think it would depend a lot on (1) how diffused those war-relevant prediction services are and (2) the distribution of societal trust in them (e.g. whether they become politicalized), which would be country/context-specific and I did not come up with useful ways to further disentangle them on a general level.