Predictive AI in Conflict Zones: Promise, Peril, and the Data We Are Still Missing

The honest scope of predictive AI in 2026

Predictive AI in conflict zones is neither the breakthrough its proponents claim nor the dystopia its critics fear. It is a maturing set of statistical techniques applied to noisy data in environments where the cost of being wrong is measured in human lives. The decade-long pattern is improvement that is uneven across geographies, model types, and decision contexts. The question worth asking in 2026 is not whether predictive AI works but where it works, where it does not, and what the remaining gaps imply for governance.

What predictive AI actually does in conflict contexts

Five families of prediction now have operational footprints.

• Event forecasting. Models such as ViEWS at Uppsala University, the JRC's Conflict Risk Index, and operational systems inside intelligence agencies estimate the probability of armed-conflict events at the country, province, or month-grid-cell level. Forecast horizons range from one month to twelve months.
• Displacement forecasting. UNHCR Project Jetson, the Danish Refugee Council's Foresight platform, and academic teams produce probabilistic estimates of cross-border and internal displacement weeks to months in advance.
• Food security forecasting. IPC and FEWS NET use machine-learning components alongside expert judgment to anticipate the next phase classification, which in turn drives anticipatory action triggers.
• Casualty estimation. Academic and NGO teams estimate direct and indirect mortality using methods that increasingly incorporate satellite-derived indicators and ML imputation of missing data.
• Tactical and operational targeting. Military and intelligence systems use predictive AI for target prioritisation. This is the most consequential and least transparent category, and the one most relevant to international humanitarian law debates.

The first four sit within humanitarian and academic governance regimes. The fifth sits within national-security regimes that have only recently begun publishing accountability frameworks.

Where the promise is real

Several wins are now well documented.

• Anticipatory action triggers funded by ML-supported forecasts have demonstrably preceded crisis onset in cyclones in Bangladesh, droughts in the Horn of Africa, and floods in West Africa. The pre-positioning effect on response times is measurable.
• Displacement nowcasting reduces the lag between event and credible-estimate in fast-moving crises, supporting earlier supply pre-positioning and earlier diplomatic engagement.
• Forecast-against-observation feedback loops catch model drift earlier than annual external evaluations, which improves model quality faster than the old paradigm allowed.
• Multi-source fusion catches anomalies humans would miss. When SAR imagery, ACLED events, and call-detail records all shift in the same direction within a short window, the joint signal flags faster than any single feed.

These are not speculative. They are documented in published evaluations from UN agencies, the World Bank, and academic teams.

Where the peril is real

The peril is also documented, and falls into four categories.

### Regime-change blindness

Models trained on the recent past underweight discontinuities. Every major conflict and displacement event of the last three years (Ukraine 2022, Sudan 2023, Gaza 2023, the regime collapse in Syria in late 2024) was underweighted by leading forecasting systems in the immediate run-up. Human analysts caught some signals; the models did not. Until models incorporate structural-shock priors better, the highest-stakes events will be the ones they miss.

### Data-sparse environments

Models trained on dense ACLED coverage perform poorly in regions where event reporting is thin. Sudan, Myanmar, parts of West Africa, and most of Central Asia have lower event-density than the model-development corpus. Confidence intervals are wider than headline numbers suggest, and the bias is systematic.

### Misuse for border closure justification

Forecasts that name specific origin and destination countries can be and have been used by host governments to pre-emptively justify border restrictions. The IASC operational guidance restricts public-facing forecasts to aggregated levels for this reason; compliance is uneven.

### Targeting and the IHL debate

Predictive AI used in military targeting raises distinct issues that the humanitarian sector cannot solve but is affected by. The ICRC and others have published positions calling for meaningful human control over lethal force decisions; the EU AI Act explicitly excludes military uses from its scope, leaving governance to other instruments. The factual record on accountability remains thin.

The data we are still missing

Three structural data gaps continue to constrain what predictive AI can do.

• Local-language event reporting at scale. ACLED and UCDP rely heavily on English-language sources, which under-represent events covered only in Arabic, French, Russian, or local languages.
• Validated training data on regime-change-class events. The historical record contains few examples, which limits model performance on exactly the events that matter most.
• Ground-truth on indirect mortality. Direct conflict deaths are countable; indirect deaths from displacement, disease, and food insecurity are estimated with methods that vary by an order of magnitude. Models that estimate humanitarian needs against an incomplete mortality ground truth inherit that uncertainty.

What good governance looks like

The agencies and academic teams operating in this space responsibly converge on a short list of practices.

• Public-facing forecasts cite uncertainty ranges, not point estimates.
• Aggregated geographies are used in public outputs; granular forecasts are restricted to operational users.
• Forecasts are paired with human-analyst review before external release.
• Methodologies are published; training-data limitations are named.
• Model performance is reported against the populations and regions the model is deployed against, not against the aggregate.
• Decision rights are documented: which forecast triggers which action, and who signs off.

These practices are not the same as a formal regulatory regime, and the gap between them and a binding international framework is the policy frontier worth watching in 2026.

What policy researchers should track

• The first operational deployments of anticipatory action triggers in conflict (not just climate) contexts.
• The evolution of EU AI Act jurisprudence on dual-use civilian-military systems.
• The IASC's next iteration of operational guidance on data responsibility.
• The publication cadence and methodology disclosure of major forecasting platforms.
• The status of human-control language in international humanitarian law discussions on autonomous and AI-supported weapons.

The 2026 picture is one of consequential capability paired with immature governance. The technical trajectory is faster than the institutional trajectory; closing that gap is the work of the next several years.

Sources and further reading

• ViEWS political violence early-warning: https://viewsforecasting.org/
• JRC INFORM Severity Index and Conflict Risk Index: https://drmkc.jrc.ec.europa.eu/inform-index
• UNHCR Project Jetson: https://www.unhcr.org/innovation/
• Danish Refugee Council Foresight: https://pro.drc.ngo/resources/news/foresight-displacement-forecasts/
• IPC and FEWS NET: https://www.ipcinfo.org/ and https://fews.net/
• ICRC position on autonomous weapon systems: https://www.icrc.org/en/document/icrc-position-autonomous-weapon-systems
• EU AI Act text: https://artificialintelligenceact.eu/