Why AI Struggles to Forecast Emerging Conflicts

Introduction

AI conflict forecasting systems are often presented as tools that can spot danger before violence begins. In some situations they do help identify escalating risks, especially in places with an established history of conflict. Yet one of the hardest problems in the field is predicting violence where little or no recent violence has occurred. The first outbreak of a new civil conflict, insurgency, ethnic clash or interstate confrontation is often exactly the event policymakers most want to prevent, but it is also the event many forecasting systems find hardest to anticipate.

This challenge matters beyond the technical performance of machine-learning models. If advanced AI is to help create a safer and more resilient civilisation, it must eventually do more than recognise patterns from past wars. It must also help institutions notice novel risks before they become disasters. The difficulty is that emerging conflicts often arise from combinations of political decisions, social tensions and unexpected shocks that do not closely resemble previous cases. Researchers increasingly describe this as one of the central limits of conflict forecasting. OUP Academic[JRC Publications]publications.jrc.ec.europa.euJRC Publications The Future of Conflict Early WarningJRC PublicationsThe Future of Conflict Early Warning - JRC Publicationsby G SCHVITZ — Machine learning has shown promise in forecasting p…

Data Scarcity Makes New Conflicts Hard to Learn

Most machine-learning systems learn by identifying recurring patterns in historical data. This works best when many examples exist.

Conflict forecasting therefore benefits from regions that have already experienced repeated violence. Previous attacks, ceasefire breakdowns, armed group activity and casualty trends create statistical signals that models can detect. Systems such as ViEWS and related forecasting platforms often use historical conflict records as some of their strongest predictors.[SSOAR]ssoar.infoViEWS: A political violence early-warning systemViEWS provides forecasts 36 months into the future for three types of political viol…[Peace Research Institute Oslo]prio.orgPeace Research Institute OsloViEWS: A political Violence Early-Warning SystemThis article presents ViEWS – a political violence early-war…

The problem is that genuinely new conflicts provide few such signals.

A region that has remained relatively peaceful for years may have:

From a machine-learning perspective, the absence of previous conflict often looks similar to ordinary stability. The model sees little evidence that distinguishes a future crisis zone from thousands of other peaceful areas.

Researchers studying conflict forecasting have repeatedly found that past violence is among the strongest predictors of future violence. That creates a structural weakness: systems become much better at forecasting conflict continuation or recurrence than conflict onset. Håvard Mueller and colleagues describe outbreaks in countries without recent violence as the “hard problem” of conflict prediction because the most informative predictor—recent conflict history—is largely absent.[OUP Academic]academic.oup.comOUP AcademicHard Problem of Prediction for Conflict Preventionby H Mueller · 2022 · Cited by 85 — We call this the hard problem of confli…

This creates an uncomfortable paradox. The conflicts most worth preventing are often those for which the available data are weakest.

Rare Events Create Statistical Blind Spots

Another difficulty is that major conflict outbreaks are rare events.

Most countries do not experience civil war in any given year. Most local districts do not suddenly become battle zones. As a result, datasets contain far more examples of non-conflict than conflict.

This imbalance creates several problems:

Researchers working on political violence forecasting have long identified rare-event prediction as a fundamental statistical challenge. Because severe violence occurs infrequently, there are relatively few examples from which algorithms can learn reliable patterns.[Benjamin E. Bagozzi]benjaminbagozzi.comBenjamin EBagozziData-based Computational Approaches to Forecasting…by PA Schrodt · 2013 · Cited by 44 — As we have stressed repeatedly, one of…

The issue becomes even more difficult when the conflict is unprecedented in form. A model trained largely on insurgencies, separatist wars or familiar patterns of civil conflict may struggle when instability emerges through a new mechanism, such as a digitally coordinated movement, a sudden state collapse, a climate-driven displacement crisis or a novel geopolitical confrontation.

Historical Patterns Can Miss Political Turning Points

Many forecasting systems are strongest when future conditions resemble past conditions.

Political crises, however, often emerge from decisions made by individuals or small groups under unusual circumstances. Leadership changes, disputed elections, assassinations, military defections, economic collapses or sudden external interventions can transform a stable political environment within weeks.

These turning points are difficult because they are not always visible in historical trend data.

For example, a model may observe:

Yet a sudden constitutional crisis or leadership struggle can rapidly change incentives for political actors.

Human analysts face the same challenge, but machine-learning systems are particularly vulnerable when their forecasts depend heavily on historical regularities. Researchers at the Alan Turing Institute’s Centre for Emerging Technology and Security note that no current AI system can reliably predict geopolitical flashpoints, partly because political events often depend on complex strategic decisions that are difficult to capture through existing datasets.[Emerging Tech & Security Center]cetas.turing.ac.ukEmerging Tech & Security CenterApplying AI to Strategic WarningThe two most promising use cases identified in this research are AI to tra…

In practical terms, models often recognise momentum better than transformation. They can detect that a conflict is worsening more easily than they can detect that a previously peaceful society is approaching a critical break.

Weak Signals Are Often Hidden in Noisy Data

Advocates of AI forecasting frequently argue that modern systems can detect subtle warning signs that humans overlook. Sometimes this is true. News reporting patterns, social-media discussions, population movements and economic changes can all provide useful information.[EURIDICE]euridice.euAn Open-Source AI Framework for Forecasting Armed ConflictAn Open-Source AI Framework for Forecasting Armed ConflictAugust 21, 2025 — 8 Sept 2025 — Surges in conflict-related news coverag…Published: August 21, 2025

The difficulty is distinguishing meaningful signals from ordinary noise.

A rise in online anger, for example, may reflect:

These possibilities can look similar in raw data.

Emerging conflicts often begin with ambiguous developments whose significance becomes obvious only in hindsight. Before violence occurs, warning indicators are frequently weak, fragmented and contradictory. The same pattern of protests, inflammatory rhetoric or local unrest may lead to civil conflict in one case and peaceful political compromise in another.

This ambiguity means that systems attempting to capture early warning signals face a difficult trade-off. Models sensitive enough to detect weak signals may generate many false alarms. Models calibrated to reduce false alarms may miss genuine emerging threats.

Forecasting Systems Often Learn Where Violence Usually Happens

Many modern conflict models incorporate spatial information. They examine where violence has occurred previously and estimate whether nearby regions face elevated risk.

This approach improves performance in many settings because conflict often spreads through geographic networks. Violence can spill across borders, follow transportation routes or move between neighbouring districts.[Taylor & Francis Online]tandfonline.comTaylor & Francis OnlineForecasting conflict in Africa with automated machine…by V D’Orazio · 2022 · Cited by 26 — The ViEWS problem is…

However, this strength can become a weakness when forecasting entirely new conflict zones.

If a region has little history of organised violence and is distant from existing conflict clusters, the model may assign low risk simply because it lacks precedent. Researchers reviewing the future of conflict forecasting note that machine-learning systems remain substantially better at identifying instability in already vulnerable regions than at anticipating conflict onset in previously peaceful societies.[JRC Publications]publications.jrc.ec.europa.euJRC Publications The Future of Conflict Early WarningJRC PublicationsThe Future of Conflict Early Warning - JRC Publicationsby G SCHVITZ — Machine learning has shown promise in forecasting p…

This creates a kind of statistical conservatism. Models become good at extrapolating from known hotspots while remaining cautious about predicting major instability in places that appear stable according to historical data.

For policymakers, this means that low predicted risk should not automatically be interpreted as evidence of safety.

New Forms of Conflict May Not Match Older Wars

Conflict forecasting systems are usually trained on datasets that classify violence according to historical categories such as civil wars, insurgencies, one-sided violence or interstate conflict.

Yet the nature of political confrontation evolves.

Emerging forms of instability may involve:

As forms of conflict change, historical datasets become less complete guides to future behaviour.

Researchers developing next-generation forecasting systems increasingly emphasise the need for models that can integrate diverse data sources and adapt to changing conflict dynamics rather than relying exclusively on historical violence patterns.[arXiv]arxiv.orgSource details in endnotes.

The broader AI-bloom question is whether future systems could become flexible enough to recognise genuinely novel threats. Current evidence suggests some improvement is possible, but robust prediction of unprecedented conflict remains an unsolved problem.

Why Better AI Alone May Not Solve the Problem

It is tempting to assume that larger models, more computing power and more data will eventually eliminate these forecasting limitations.

There are reasons for caution.

Some forecasting failures arise not from insufficient computational capability but from genuine uncertainty in the world itself.

Human actors change behaviour in response to new information. Political leaders conceal intentions. Secret negotiations occur outside observable datasets. Chance events alter incentives. Some conflicts emerge from unique combinations of circumstances that have never previously occurred.

Even highly capable future AI systems may therefore face irreducible uncertainty.

Several researchers argue that conflict forecasting should be understood as probabilistic risk assessment rather than prediction in the everyday sense of knowing what will happen. Modern systems can improve estimates of likelihood, but they cannot remove uncertainty from inherently contingent political events. White Rose Research Online ScienceDirect This distinction matters because unrealistic expectations can create disappointment and mistrust. A forecast that correctly identifies a regi[sciencedirect.com]sciencedirect.comA review and comparison of conflict early warning systemsby EG Rød · 2024 · Cited by 53 — A conflict early warning system (CEWS) is a ris… on as having elevated risk may still fail to predict the exact outbreak, timing or form of violence.

What This Means for Prevention Strategies

The limits of forecasting emerging conflicts have important implications for how early-warning systems should be used.

First, forecasts are often more valuable as tools for prioritising attention than as definitive predictions. A low-confidence warning may still justify closer monitoring, diplomatic engagement or contingency planning.[VIEWS]viewsforecasting.orgVIEWSViolence Early-Warning SystemThe Violence & Impacts Early-Warning System. We deliver scalable, data-driven tools to empower decision…

Second, quantitative models work best when combined with human expertise. Local researchers, journalists, civil-society organisations and regional specialists often possess contextual knowledge that does not appear in structured datasets. Their insights can help interpret ambiguous signals that algorithms alone may misread.

Third, prevention systems may need to focus less on predicting a single outbreak and more on building resilience against a range of plausible risks. If truly novel conflicts remain difficult to forecast, strengthening institutions, improving governance, supporting peaceful dispute resolution and reducing vulnerability to shocks may provide protection even when forecasts are uncertain.

Within the larger vision of AI supporting humanity’s long-term flourishing, this is a useful reminder of both the promise and the limits of prediction. Advanced AI may eventually become far better at detecting weak signals, integrating global information and modelling complex political systems. Yet preventing first-time conflicts is unlikely to become a simple technical problem. The challenge is not merely processing more data. It is understanding societies well enough to recognise when a peaceful future is beginning to fracture before the fracture becomes visible to everyone else.[Emerging Tech & Security Center]cetas.turing.ac.ukEmerging Tech & Security CenterApplying AI to Strategic WarningThe two most promising use cases identified in this research are AI to tra…[JRC Publications]publications.jrc.ec.europa.euJRC Publications The Future of Conflict Early WarningJRC PublicationsThe Future of Conflict Early Warning - JRC Publicationsby G SCHVITZ — Machine learning has shown promise in forecasting p…

Amazon book picks

Further Reading

Books and field guides related to Why AI Struggles to Forecast Emerging Conflicts. Use these as the next step if you want deeper reading beyond the article.

Book

The Signal and the Noise

By Nate Silver

Explains why rare events are difficult to predict.

See on Amazon

Book

The Black Swan

By Nassim Nicholas Taleb

Rating: 4.0/5 from 25 Google Books ratings

Focuses on unexpected events that evade prediction.

See on Amazon

Book

Thinking in Bets

By Annie Duke

Useful for reasoning about uncertain conflict forecasts.

See on Amazon

Book

Superforecasting

By Philip E. Tetlock, Dan Gardner

Directly relevant to prediction limits and emerging risks.

See on Amazon

Browse more on Amazon:The Signal and the NoiseThe Black SwanThinking in Bets

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Example marketplace items related to this page. Use the search link to explore similar finds on eBay.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Example eBay listing

WORLD MAP POSTER LAMINATED EDUCATIONAL SCHOOL WALL ART PRINT A5 A4 A3 A2 A1

SearcheBay.co.uk: world map poster

Browse similar oneBay.co.uk

Example eBay listing

World Map Maps Educational Maxi Poster Print 61x91.5cm | 24x36 inches

SearcheBay.co.uk: world map poster

Browse similar oneBay.co.uk

Example eBay listing

Gleneagles Scratch Off World Map Poster With Flags - White, Extra Large 60x85 cm

SearcheBay.co.uk: world map poster

Browse similar oneBay.co.uk

Example eBay listing

World Map Atlas Detailed Large Poster Art Print Gift - A0 A1 A2 A3 A4 A5 + More

SearcheBay.co.uk: world map poster

Browse similar oneBay.co.uk

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: academic.oup.com
Link:https://academic.oup.com/jeea/article/20/6/2440/6574413

2. Source: ssoar.info
Link:https://www.ssoar.info/ssoar/bitstream/document/62201/1/ssoar-jpeaceresearch-2019-2-hegre_et_al-ViEWS_A_political_violence_early-warning.pdf

3. Source: sciencedirect.com
Link:https://www.sciencedirect.com/science/article/pii/S0169207023000018

4. Source: benjaminbagozzi.com
Title: Benjamin E
Link:https://www.benjaminbagozzi.com/uploads/1/2/5/7/12579534/data-based-computational-approahes-to-forecasting-political-violence.pdf

5. Source: arxiv.org
Link:https://arxiv.org/abs/1903.00604

6. Source: euridice.eu
Title: An Open-Source AI Framework for Forecasting Armed Conflict
Link:https://euridice.eu/wp-content/uploads/2025/09/Beyond-Closed-Doors-An-Open-Source-AI-Framework-for-Forecasting-Armed-Conflict-1.pdf

Published: August 21, 2025

7. Source: arxiv.org
Link:https://arxiv.org/html/2506.14817v1

8. Source: viewsforecasting.org
Link:https://viewsforecasting.org/

9. Source: nature.com
Link:https://www.nature.com/articles/s41598-025-11812-2

10. Source: academic.oup.com
Link:https://academic.oup.com/jrsssa/advance-article/doi/10.1093/jrsssa/qnag039/8539545

11. Source: publications.jrc.ec.europa.eu
Title: JRC Publications The Future of Conflict Early Warning
Link:https://publications.jrc.ec.europa.eu/repository/bitstream/JRC143004/JRC143004_01.pdf

12. Source: cetas.turing.ac.uk
Link:https://cetas.turing.ac.uk/publications/applying-ai-strategic-warning

13. Source: prio.org
Link:https://www.prio.org/publications/11231

14. Source: tandfonline.com
Link:https://www.tandfonline.com/doi/full/10.1080/03050629.2022.2017290

15. Source: tandfonline.com
Link:https://www.tandfonline.com/doi/abs/10.1080/03050629.2022.2031182

16. Source: viewsforecasting.org
Link:https://viewsforecasting.org/news/views-featured-in-the-economist-on-ai-and-conflict-prediction/

17. Source: viewsforecasting.org
Title: Known Issues | VIEWS
Link:https://viewsforecasting.org/early-warning-system/known-issues/

18. Source: tandfonline.com
Title: Vi EWS: A political violence early-warning
Link:https://www.tandfonline.com/doi/full/10.1080/03050629.2022.2090933

Additional References

19. Source: trendsresearch.org
Title: the impact of ai and machine learning on conflict prevention
Link:https://trendsresearch.org/insight/the-impact-of-ai-and-machine-learning-on-conflict-prevention/?srsltid=AfmBOoost7eH6QSh5RvUEGABfLb34Kha3V5RmJgL1pJyShpwFd7ubMtA

Published: May 2025

20. Source: cordis.europa.eu
Title: CORDISUsing machine learning to identify political violence
Link:https://cordis.europa.eu/article/id/443344-using-machine-learning-to-identify-political-violence-and-anticipate-conflict

21. Source: visionofhumanity.org
Title: predicting civil conflict can machine learning tell us
Link:https://www.visionofhumanity.org/predicting-civil-conflict-can-machine-learning-tell-us/

22. Source: trendsresearch.org
Title: The Impact of AI and Machine Learning on Conflict
Link:https://trendsresearch.org/insight/the-impact-of-ai-and-machine-learning-on-conflict-prevention/?srsltid=AfmBOorCUEP5dv-cBNyycT5Yh5L92jovEPVfvP8wrpx8mPYIqBeM4QZp

Published: May 2, 2025

23. Source: royalsocietypublishing.org
Title: Data driven conflict classification exposes weak
Link:https://royalsocietypublishing.org/rsos/article/12/12/250897/366130/Data-driven-conflict-classification-exposes-weak

24. Source: oecd-opsi.org
Title: Observatory of Public Sector Innovation Vi EWS
Link:https://oecd-opsi.org/innovations/views-the-political-violence-early-warning-system/

25. Source: economist.com
Title: ai models are being used to predict conflict
Link:https://www.economist.com/science-and-technology/2026/05/13/ai-models-are-being-used-to-predict-conflict

Published: May 2026

26. Source: cambridge.org
Link:https://www.cambridge.org/core/journals/data-and-policy/article/promise-of-machine-learning-in-violent-conflict-forecasting/40D559ADA18FF7308915B08956B4E8F3

27. Source: researchgate.net
Link:https://www.researchgate.net/publication/383579991_The_promise_of_machine_learning_in_violent_conflict_forecasting

28. Source: tandfonline.com
Link:https://www.tandfonline.com/doi/abs/10.1080/03050629.2022.1993209