Introduction: from uncertainty to prediction

Modern conflict has always involved uncertainty. Political leaders, military commanders, intelligence agencies, diplomats, and non-state actors act under conditions of incomplete information, as they do not fully know what the other side wants, how far it is willing to go, whether threats are credible, or whether apparent restraint hides preparation for escalation.

War and peace are therefore shaped not only by weapons, interests, and ideology, but also by expectations [8]. Artificial intelligence changes this landscape because it promises to transform uncertainty into calculation.

Predictive systems can process satellite images, social media signals, economic flows, military deployments, cyber activity, diplomatic language, and behavioral patterns in order to forecast possible crises. In principle, this could help prevent violence, as early-warning systems may detect escalation before human observers do. Diplomats may test likely reactions before making a move and humanitarian workers may anticipate displacement. However, the same predictive power can also intensify conflict. When actors believe they can calculate the future, they may act sooner, harder, and with greater confidence. The danger is not simply that AI may be wrong. The deeper danger is that AI predictions may become part of the conflict itself [10][13][14][15].

Recent strategic-stability literature already warns that AI may compress decision-making time, increase risks of machine-speed escalation, create new forms of miscalculation, and weaken traditional human “firebreaks” against escalation [10][14]. SIPRI’s work on AI and strategic stability similarly highlights the relevance of AI for command and control, strategic offence and defence, and nuclear deterrence systems. This means that predictive AI should not be treated as a neutral tool added to an unchanged strategic environment. It changes the structure of that environment.

In this sense, three frameworks help clarify what is at stake: Jean-Pierre Dupuy’s philosophy of catastrophe and deterrence, game theory, and the logic of common knowledge. What each reveals, in different but complementary ways, is that conflict escalation is not only a matter of material capability—it is also a matter of how actors imagine the future, how they model each other, and what they believe the other side believes [11].

Dupuy: enlightened doomsaying and the paradox of deterrence

Jean-Pierre Dupuy’s work is especially useful because he does not treat catastrophe merely as an unfortunate event that may or may not happen. He asks how the imagined future acts upon the present [4]. His idea of “enlightened doomsaying” is not simple pessimism, but it is rather a philosophical attempt to make catastrophe sufficiently real in the present so that actors take action to prevent it. Dupuy’s central problem is framed as fundamentally paradoxical—if a catastrophic warning succeeds, the catastrophe does not occur, but if it does not occur, people may later conclude that the warning was exaggerated. In other words, the successful warning cancels the evidence that would have proved it necessary [3][4].

This is highly relevant to AI-driven conflict prediction. Suppose an AI system predicts a high probability of military escalation. Decision-makers may respond by mobilizing forces, issuing deterrent threats, launching pre-emptive cyber operations, or placing systems on alert. If escalation does not occur, they may claim that the AI “prevented” the crisis. But another interpretation is also possible: the AI prediction may itself have intensified the crisis, creating the very danger it claimed to anticipate [10][15].

Predictive systems therefore operate inside a circular structure: they do not merely describe the future, but they can help produce the future that later appears to justify them.

Dupuy’s thought also helps explain the paradox of deterrence. Deterrence depends on the credibility of a threat that one hopes never to carry out. Nuclear deterrence is the clearest case: the threat must be believable enough to prevent aggression, but if it fails, the result may be catastrophic. Dupuy’s writings on MAD (Mutually Assured destruction) emphasize this paradoxical structure of deterrence—deterrence works only if the catastrophic possibility is taken seriously, but at the same time its success consists in making that catastrophe not happen [3][4].

AI complicates deterrence because it can appear to make threats more precise, more credible, and more rapidly executable. A state may believe that AI allows it to identify enemy vulnerabilities, predict enemy responses, and strike before the enemy can adapt. But this can produce what might be called algorithmic overconfidence. If each side believes its models reveal the other side’s future behavior, conflict becomes less a matter of diplomacy and more a race to act before predicted risks materialize. The future could be no longer awaited—it could be operationalized [10][14][15].

Dupuy’s notion of “metaphysical MAD” is also useful here. Classical MAD — mutually assured destruction — is a strategic doctrine. Metaphysical MAD is deeper: it concerns the structure of a world in which humanity must live under the shadow of a catastrophe that must be treated as real precisely in order not to occur [3][4]. AI intensifies this condition because it gives catastrophe a technical interface. The future appears not as myth, prophecy, or political intuition, but as dashboard, probability, simulation, and alert level. The problem is that calculation may weaken the humility that catastrophe requires. Dupuy’s lesson is not simply “predict disaster”. It is that catastrophe must be imagined in a way that interrupts normal strategic calculation. AI, by contrast, may normalize catastrophic futures by turning them into manageable scenarios [3][12]. The risk is that nuclear escalation, civil war, ethnic violence, or regional collapse becomes one option among others in a decision-support system.

Game theory: prediction, strategic interaction, and escalation traps

Game theory studies situations in which the outcome of one actor’s decision depends on the decisions of others. Conflict is therefore a natural field for game-theoretic analysis. A state does not simply choose between war and peace in isolation, but it chooses under the expectation of how the adversary will respond. At the same time, the adversary does the same, therefore each side tries to anticipate the anticipation of the other [11].

AI strengthens this recursive structure, as it can model likely enemy behavior, simulate bargaining scenarios, estimate resolve, and recommend optimal moves. This may improve decision-making when the model is accurate and when political leaders understand its limits.  But it may also produce escalation traps [10][15].

The first trap is the “trap of preemptive action”. If an AI system predicts that the adversary is likely to attack, the rational recommendation may be to move first. But at the same time, if the adversary has a similar system, it may interpret this movement as confirmation of hostile intent. Each side then acts defensively in its own eyes, while appearing aggressive to the other.

This is the classical security dilemma, accelerated by predictive analytics [9][10]. James Fearon argues that war occurs despite its costs due to rational failures in bargaining, driven by information asymmetry, commitment problems, and issue indivisibility. War is therefore a voluntary choice made when states cannot reach a mutually beneficial peaceful agreement, challenging the notion that conflicts are solely due to irrationality [6].

The second trap is the “commitment trap”. In many conflicts, actors try to make their threats credible by tying their own hands: public red lines, military deployments, alliance commitments, or automatic response systems. AI may strengthen such commitments by integrating prediction into command structures. But the more automatic and data-driven the response becomes, the less room remains for reinterpretation, delay, ambiguity, and face-saving compromise. In crises, these “soft” elements are often not weaknesses, but they are strategically used as de-escalatory resources.

The third trap is the “optimization trap”. Game-theoretic reasoning often assumes that actors maximize expected utility. AI systems are especially powerful optimization machines. But conflict is not a chessboard with stable rules and clear payoffs. Human actors care about honor, fear, humiliation, domestic legitimacy, symbolic recognition, revenge, historical memory, and ideological narratives. A system that optimizes for “advantage” may misread the political meaning of restraint or escalation. It may recommend moves that are formally rational but politically disastrous [2].

The fourth trap is the speed trap. AI can compress decision-making timelines. In military contexts, speed is often treated as an advantage. But in diplomacy and crisis management, speed can be dangerous. De-escalation often requires delay: time to verify information, communicate intentions, consult allies, create off-ramps, and allow public rhetoric to cool. Several analyses of AI and strategic stability warn that machine-speed warfare, miscalculation, and the erosion of human firebreaks may increase escalation risk [7][12][14].

The central game-theoretic point is this: AI does not simply help one player make better choices. It changes the game for all players. If one actor uses predictive AI, others must respond to the fact that they are being modeled. They may then attempt to deceive the model, manipulate its inputs, or behave unpredictably in order to avoid being strategically captured. This can produce an arms race not only in weapons, but in prediction, deception, and counter-prediction.

This is why AI may both stabilize and destabilize conflict. It can stabilize when it improves early warning, reduces uncertainty, detects accidental escalation, and identifies mutually beneficial settlements. It destabilizes when it increases confidence in pre-emption, encourages rapid action, automates threat perception, or creates incentives to manipulate the adversary’s model.

Common knowledge and epistemic modal logic: what each side knows that the other side knows

Common knowledge is one of the most important but least publicly understood concepts in conflict analysis. A fact is not merely known when one actor knows it. It becomes common knowledge when everyone knows it, everyone knows that everyone knows it, everyone knows that everyone knows that everyone knows it, and so on. In conflict, common knowledge matters because coordination depends on shared expectations. A ceasefire may fail not because both sides reject peace, but because neither side trusts that the other side knows that restraint is expected. A deterrent threat may fail not because it is weak, but because it is not commonly understood as credible. A diplomatic signal may fail not because it is absent, but because its interpretation is not mutually recognized.

Leaders do not act only on facts. They act on beliefs about others’ beliefs. AI enters this structure in several ways.

First, AI may create asymmetric knowledge. One side may possess predictive tools that the other side does not. This can produce strategic advantage, but also suspicion. If the weaker side believes the stronger side can predict its moves, it may act unpredictably, hide information, or escalate early before its options disappear. Second, AI may create false common knowledge. If both sides rely on similar data sources, similar models, or shared public indicators, they may converge on the same mistaken interpretation. For example, troop movements, cyber anomalies, or social media signals may be interpreted by multiple systems as preparation for attack. Once this interpretation becomes mutually expected, it can become performative, as each side believes that the other side believes escalation is coming, and therefore prepares for escalation [1][10]. Third, AI may weaken the role of constructive ambiguity. In diplomacy, ambiguity is not always a defect. Sometimes peace depends on allowing different actors to interpret the same agreement in slightly different ways. Common knowledge is necessary for coordination, but total transparency is not always necessary for peace. AI systems, however, often seek to reduce ambiguity, as they primary classify, rank, score, and predict. This may be useful in some domains, but conflict de-escalation often requires symbolic flexibility [8].

The classic “coordinated attack” problem in common knowledge theory shows that even when actors exchange messages, perfect coordination may remain impossible if they cannot be certain that the message was received, that receipt was known, and so on. This problem has direct relevance to crisis communication. De-escalation does not require only sending signals; it requires creating a sufficiently stable shared belief that the signals have been received and understood [5].

AI may help here. It can monitor whether signals are being interpreted correctly, identify miscommunication, and model likely misunderstandings. But it may also harm the process if it encourages actors to treat probabilistic inference as equivalent to mutual understanding. Prediction is not the same as recognition. An actor may correctly predict another actor’s next move while still misunderstanding its reasons, fears, constraints, and symbolic universe [8][10].

How algorithmic prediction escalates conflict

Algorithmic prediction can escalate conflict through several mechanisms. The first is “anticipatory action”. If the future is predicted as dangerous, actors may act now to prevent being trapped later. This is the logic of pre-emption. The problem is that pre-emption often appears defensive to the actor who undertakes it and offensive to the actor who observes it [9].

The second is “feedback amplification”. AI systems learn from signals, but in conflict, signals are often reactions to previous signals. If one side raises readiness because an AI predicts danger, the other side’s AI may interpret that readiness as evidence of danger. The system then generates a loop: (1) prediction produces action; (2) action produces confirmation; (3) confirmation produces further prediction [4][10][15].

The third is “loss of political context”. AI may detect patterns without understanding their political meaning. A military exercise, domestic speech, symbolic event, or alliance meeting may be read as escalation. Human analysts can also misread such signals, of course. But AI can scale misinterpretation and give it the authority of technical objectivity [8][12].

Another factor is “moral distancing”. Prediction can make escalation appear impersonal. A leader may say: “The system indicates that escalation is inevitable”, rather than “We chose escalation.” This shifts responsibility from judgment to calculation. The political danger is that AI becomes a mechanism for laundering decisions: human actors remain responsible, but the machine provides the language of necessity [4][12].

How algorithmic prediction can support de-escalation

First, AI can improve “early warning”. It may identify risks before they become visible to human institutions. Second, AI can support “scenario testing”. Policymakers can simulate how different actions may be perceived by adversaries, allies, domestic populations, and international organizations. Used carefully, this may help identify moves that reduce fear rather than intensify it. Third, AI can help detect “misperception”. If one side interprets a signal as aggressive, AI-assisted analysis may compare this interpretation with alternative explanations. This could create space for verification before retaliation. Fourth, AI can support “narrative modelling”. De-escalation depends not only on what is said, but on how, when, by whom, and through which channel. AI may help identify language that is firm enough to preserve credibility but not so humiliating that it forces the other side to escalate.

But these benefits depend on a crucial condition and it is that AI must remain advisory rather than sovereign. It should widen judgment, not replace it. It should make possible futures visible, not convert them into destiny.

The central paradox: prediction changes the predicted future

The deepest issue is that conflict prediction is reflexive. Predicting an earthquake does not usually change the earthquake, but predicting a bank run, however, may cause one. Conflict is closer to the bank run than to the earthquake, as the prediction enters the field it predicts. This is where Dupuy, game theory, and epistemic logic converge in the context of the current analysis.

From Dupuy, we learn that catastrophic futures can act upon the present. A predicted catastrophe may prevent itself, but it may also produce the actions that make it more likely. The future is not simply ahead of us, but it is rather already active in our present decisions.

From game theory, we learn that each actor’s rational choice depends on expectations about other actors’ choices. AI changes those expectations by making prediction faster, more systematic, and more operational. The result is a new strategic condition: algorithmic reflexivity. This means that actors increasingly act not only in response to the adversary, but in response to predictions of the adversary, predictions of the adversary’s predictions, and predictions of how their own behavior will be predicted. Conflict becomes a struggle over the future as modeled by machines.

Practical implications for conflict analysis

For conflict analysts, the rise of AI prediction requires several methodological adjustments that go beyond technical evaluation. The central question is not only whether a given prediction is accurate, but how that prediction will shape behavior once it is believed—a distinction that matters especially when the subject of forecasting is a potential adversary. Equally important is the difference between forecasting and strategic signaling: where forecasting asks what is likely to happen, signaling is a communicative act that conveys intention and resolve. Treating signals as mere data points risks misreading the political grammar of crisis. Analysts should also consider whether AI systems increase or decrease common knowledge: a prediction available only to one side may be destabilizing, whereas a mutually recognized early warning may help de-escalate. This points to a broader concern about speed. In crisis situations, faster information is not always better information, and systems that compress decision time may compress the quality of judgment along with it. Finally, the question of political responsibility cannot be set aside. Decision-makers must not be allowed to shelter behind predictive systems—“the algorithm recommended it” is not a legitimate substitute for strategic, legal, or moral accountability.

Conclusion: against the illusion of a fully calculable future

AI will not abolish uncertainty from conflict, but it will rather reorganize uncertainty. It may reduce some unknowns while creating others: uncertainty about the model, uncertainty about how predictions are interpreted, and uncertainty about whether machine-generated forecasts are becoming self-fulfilling. The central danger is not simply that AI may miscalculate the future. The greater danger is that political actors may begin to treat the calculated future as inevitable. Once this happens, prediction becomes fatalism, and strategy becomes obedience to a model. Dupuy’s enlightened doomsaying offers a different lesson. The point of imagining catastrophe is not to surrender to it, but to prevent it. Game theory reminds us that no actor controls the game alone. Epistemic logic reminds us that peace depends not only on information, but on shared understanding.

Integrating AI into conflict analysis therefore demands caution, humility, and institutional restraint. The future of conflict will not be determined only by weapons, data, or algorithms. It will also be determined by how human beings interpret the futures that machines place before them. The decisive question is therefore not whether AI can calculate the future. The question is whether we can prevent calculated futures from becoming self-fulfilling catastrophes.

---

1. [1] Aumann, R. J. (1976). Agreeing to disagree. The Annals of Statistics, 4(6), 1236–1239.
2. [2] Brams, S. J. (1994). Theory of moves. Cambridge University Press.
3. [3] Dupuy, J.-P. (2012). The precautionary principle and enlightened doomsaying. Revue de Métaphysique et de Morale, (4), 577–592. http://www.jstor.org/stable/23353278
4. [4] Dupuy, J.-P. (2022). How to think about catastrophe: Toward a theory of enlightened doomsaying (M. R. Anspach, Trans.). Michigan State University Press.
5. [5] Fagin, R., Halpern, J. Y., Moses, Y., & Vardi, M. Y. (1995). Reasoning about knowledge. MIT Press.
6. [6] Fearon, J. D. (1995). Rationalist explanations for war. International Organization, 49(3), 379–414.
7. [7] Horowitz, M. C. (2026). Artificial intelligence and the future of strategic stability. Texas National Security Review.
8. [8] Jervis, R. (1976). Perception and misperception in international politics. Princeton University Press.
9. [9] Jervis, R. (1978). Cooperation under the security dilemma. World Politics, 30(2), 167–214.
10. [10] Johnson, J. S. (2020). Artificial intelligence: A threat to strategic stability. Strategic Studies Quarterly, 14(1), 16–39.
11. [11] Myerson, R. B. (1991). Game Theory: Analysis of Conflict. Harvard University Press. https://doi.org/10.2307/j.ctvjsf522
12. [12] Scharre, P. (2018). Army of none: Autonomous weapons and the future of war. W. W. Norton.
13. [13] SIPRI. (2019). The impact of artificial intelligence on strategic stability and nuclear risk: Volume I: Euro-Atlantic perspectives. Stockholm International Peace Research Institute.
14. [14] SIPRI. (2020). Artificial intelligence, strategic stability and nuclear risk. Stockholm International Peace Research Institute.
15. [15] Williams, H. (2024). Algorithmic stability: How AI could shape the future of deterrence. Center for Strategic and International Studies.