From Black Box to Watchtower: Governing AI in the Age of Conflict

When a militarized dispute can escalate within hours, the tools we use to anticipate cannot afford to be opaque. Artificial intelligence (AI) is already shaping how governments assess threats, monitor flashpoints and weigh the costs of intervention. Yet the systems doing this work often operate as black boxes, producing probabilities without explanation and conclusions without accountability. That gap between predictive power and institutional trust is not a technical footnote. It is a governance crisis.

The Opacity Problem

For decades, high-performing models, including Support Vector Machines and deep neural networks, have demonstrated impressive accuracy in forecasting conflict. But accuracy alone is insufficient when the stakes are war and peace. A prediction score that cannot be interrogated cannot be challenged, audited or owned. When policymakers act on outputs they do not understand, they are not exercising judgment; they are outsourcing it.

The result is what might be called algorithmic superstition: a quiet deference to machine outputs, dressed up as evidence-based decision-making. In diplomatic contexts, this is not merely intellectually unsatisfying. It is dangerous. Misread signals, unchallenged assumptions and opaque escalation pathways have historically contributed to catastrophic miscalculation. Adding an unaccountable AI layer does not reduce this risk; it compounds it.

Rough sets and the value of legible uncertainty

One underappreciated alternative is Rough Set Theory, a framework that treats ambiguity not as a flaw to be engineered away but as a signal worth preserving. Rather than forcing geopolitical complexity into clean probabilistic outputs, Rough Sets organize knowledge into zones of certainty, possibility and indeterminacy. The boundary region, where conflict is neither clearly likely nor clearly avoidable, is not the model’s weakness; it is its most important output. Complementing this, fuzzy logic offers a way to represent gradations of truth, capturing the reality that geopolitical conditions are rarely binary and instead exist along continua such as “high tension,” “moderate instability,” or “low risk.” While rough sets delineate the structure of uncertainty, fuzzy systems quantify its degree, assigning interpretable membership values that reflect partial belonging rather than rigid classification.

Expanded rationality without governance is not a solution. It is a new category of risk.

For diplomats and security analysts, this combination is strategically powerful. Rough sets pinpoint where uncertainty is concentrated, directing attention and resources to the most volatile cases, while fuzzy logic refines these insights by quantifying how strongly specific conditions contribute to risk. Together, they enable earlier, more targeted intervention without sacrificing interpretability. Crucially, both approaches produce linguistic, rule-based outputs in the form of traceable statements linking observable conditions to anticipated outcomes, rather than the opaque numerical weights of conventional neural networks. This creates an audit trail. Policymakers can verify what the system relies on, challenge assumptions that seem flawed, and take genuine ownership of the decisions that follow.

From bounded rationality to governed rationality

The political theorist Herbert Simon described human decision-making as bounded, constrained by limited information, cognitive capacity, and time. AI systems expand those bounds, acting as what we might call rationality multipliers. They process more data, identify more patterns and model more scenarios than any human analyst could alone.

But expanded rationality without governance is not a solution. It is a new category of risk. Bias embedded in training data is amplified at scale. Models trained on historical patterns may misinterpret novel configurations. Hallucinated confidence can masquerade as rigorous analysis. The challenge, therefore, is not simply to build more powerful models that improve our ability to anticipate conflict. It is to govern the expansion of rationality itself, ensuring that AI extends human judgment. Human judgment should remain in the loop in high-stakes conflicts.

This means embedding transparency, traceability, and human oversight into AI systems by design, not as afterthoughts. It means evaluating systems not only for predictive accuracy but also for explainability, auditability, and alignment with legal and ethical standards. And it means keeping human judgment structurally central to high-stakes decisions, not as a formality but as a genuine safeguard.

The multilateral imperative

No single state can set the norms for responsible AI use in security contexts, nor should it. Conflict prevention is inherently transnational. The signals that matter, such as refugee flows, arms transfers, economic shocks and political violence, cross borders, and so must the frameworks that govern the systems reading them.

Multilateral institutions play a critical yet largely unfulfilled role. They can facilitate data-sharing agreements among governments that hoard intelligence for competitive advantage. They can promote interoperability across national early warning systems. They can reduce technological asymmetries that risk turning AI-powered conflict forecasting into a tool of great-power dominance rather than collective security.

Sovereignty and collective responsibility will inevitably pull in different directions. 

This requires creative governance architecture: trusted data spaces with defined access rules, independent international auditing bodies with real authority to assess high-risk AI systems, and shared transparency standards that do not require states to disclose sensitive capabilities. The goal is a minimum viable layer of accountability, sufficient to prevent misinterpretation, reduce strategic miscalculation and build the cross-border trust that effective conflict prevention requires.

Sovereignty and collective responsibility will inevitably pull in different directions. That tension cannot be resolved, but it can be managed through institutions designed for that purpose.

Hybrid systems, hybrid trust

The most effective early warning systems will not be purely interpretable or purely high-performance. They will be hybrids, combining the predictive power of complex models with the legibility of transparent approaches such as Rough Sets and Neuro-Fuzzy logic. High-performing models identify risk at scale; interpretable models explain it in terms that decision-makers can act on. Neither alone is sufficient. Together, they constitute more than a forecasting tool: a system capable of building trust between machines and the humans who must ultimately answer for what those machines recommend.

That trust is not a luxury. In conflict prevention, it is the precondition for everything else.

The language of peace

The ambition of AI in conflict management should not be to eliminate uncertainty, which is impossible, but to make it legible, contestable and governable. A system that tells a diplomat “conflict is 70% likely” has done something. A system that explains why, where the evidence is weakest, and what assumptions are driving the conclusion has done something far more useful.

The language of peace has always required precision, nuance, and the courage to act on incomplete information. AI can enhance that capacity, but only if we insist that it speak in terms we can understand, challenge and ultimately take responsibility for. Prediction without accountability is merely a more sophisticated way of not knowing. Governed intelligence is something entirely different.