Strategic warning – the early detection of conflict, instability, or adversarial intent – is a pillar of national security. But today’s intelligence analysts are working with outdated infrastructure, brittle data ecosystems, and escalating cognitive load. The result? Strategic surprise remains a persistent risk.
A recent research collaboration between The Alan Turing Institute’s Centre for Emerging Technology and Security (CETaS) and the Special Competitive Studies Project (SCSP) confronts this problem head-on. Their 2025 report, Applying AI to Strategic Warning, explores whether AI can help analysts anticipate key geopolitical events, what it would take to build such a capability, and the potential payoffs – and costs – of doing so.
The conclusion is measured but clear: AI is not yet capable of predicting geopolitical flashpoints with high reliability. But if properly developed and applied, AI can significantly improve early warning by:
- monitoring conflict risk indicators more accurately,
- generating scenarios immediately after a triggering event,
- helping analysts focus on under-monitored regions,
- and enabling faster, more resilient decision-making.
Crucially, this isn’t a narrow question of better algorithms. It’s about redesigning the data infrastructure and analytical ecosystems underpinning the national security enterprise.
The real value of AI in strategic warning
The report emphasises that while current AI systems can’t replace human judgment, they offer transformative potential as decision support tools. Today’s analysts face an overwhelming data burden. Even in heavily monitored regions like the Taiwan Strait or eastern Ukraine, relevant signals are scattered across multiple unstructured sources – satellite feeds, social media, economic indicators, open-source intelligence (OSINT), and classified streams.
AI systems can help by:
- Detecting anomalies earlier across vast data streams (e.g. troop movements, sentiment shifts, economic instability),
- Triangulating across multiple models to reduce individual bias or blind spots,
- Synthesising fast-moving developments into scenario forecasts that can be challenged, not blindly accepted,
- Freeing up analysts to focus on strategic interpretation, rather than data wrangling.
However, this potential is constrained by two deep challenges: the scarcity of structured geopolitical data for training AI, and the difficulty of modelling human decision-making, especially when decisions are impulsive, deceptive, or context-dependent.
Current limitations: data, humans, and noise
Strategic warning isn’t weather forecasting. Geopolitical events are rare, noisy, and often driven by opaque decision-making under uncertainty. The report outlines two core challenges.
1. Data scarcity, inconsistency, and fusion
Unlike meteorological data, there’s no consistent, high-quality dataset of past conflicts, tipping points, or strategic shocks. Data is patchy, non-standardised, and often missing entirely – especially in regions with low surveillance coverage. Existing datasets don’t sufficiently account for:
- micro-level stabilisation dynamics (e.g. subsidies, military doctrine),
- behavioural signals (grievances, biases, decision-maker intent),
- rare triggers like sudden protests or lone-actor violence.
Efforts to integrate such data – especially across classified and unclassified sources – face major traceability, validation, and governance issues.
2. Modelling human decision-making
Even with high-fidelity data, capturing the logic of political decisions is profoundly hard. Leaders’ choices may be:
- impulsive (e.g. spur-of-the-moment escalations),
- deceptive (e.g. misinformation campaigns),
- or made by unexpected actors (e.g. the fruit seller whose self-immolation sparked the Arab Spring).
- AI can help map probable risk zones, but predicting exact triggers or tipping points remains elusive.
Three phases toward strategic AI readiness
Rather than aiming immediately for an all-encompassing simulation, the report proposes a three-phase strategy grounded in pragmatism and organisational realism.
Phase 1: data foundations
The first step isn’t more AI – it’s better data. This means creating a consistent, high-resolution repository of conflict risk indicators, triggers, and stabilisers. It includes:
- measured data (e.g. geospatial, economic, troop movements),
- non-traditional sources (e.g. mobile app metadata, leaks, OSINT),
- mental models (e.g. leader beliefs, grievances, regional sentiment).
This phase also includes hard work on data validation, anti-poisoning techniques, metadata tagging, and overcoming bureaucratic friction around data sharing.
Success here means analysts can work with a much richer baseline of labelled events – essential for training models and simulating plausible futures.
Phase 2: best-in-class model ecosystem
Rather than building one grand model, Phase 2 focuses on cultivating a suite of complementary models:
- Some built in-house, some from trusted vendors.
- Some trained on open data, others on sensitive or classified sources.
- Some global, others hyper-local and fine-tuned.
This ecosystem provides multiple, triangulated outputs to human analysts. Crucially, these are inputs, not final judgments. Human-machine teaming remains central. Models can help analysts detect anomalies, validate hypotheses, or explore edge-case scenarios – but interpretation, context, and strategic framing stay human-led.
Phase 3: integrated AI-based simulation
The final phase envisions an agentic AI simulation platform capable of fusing geopolitical, economic, social, and behavioural data into large-scale scenario analysis. This moonshot capability would:
- simulate thousands of possible futures based on conflict indicators,
- test the impact of policy choices on potential outcomes,
- help understand systemic interdependencies (e.g. how inflation, migration, and military posture interact).
But this demands massive compute, synthetic data generation, new forms of explainability, and breakthroughs in causal inference and decision science.
Even then, the goal isn’t perfect prediction. It’s helping analysts and policymakers understand what futures are possible – and where intervention could shift the odds.
Policy and cost landscape
The report is clear-eyed about the financial and political implications of building such a capability. Key issues include:
• cost: Depending on scope, a full implementation could resemble the establishment of a new combatant command in terms of budget and headcount. Building from scratch may cost billions; even commercial partnerships require substantial investment in data rights, security, and infrastructure.
• AI sovereignty and alliances: Should this be pursued nationally, or collaboratively with allies like the US, Australia, or EU states? Multilateral efforts reduce duplication but introduce interoperability, trust, and regulatory complexity.
• industry collaboration: Partnering with vendors may accelerate delivery but raises questions of control, auditability, and long-term autonomy. Open standards and modular architecture will be essential.
There’s also a less tangible – but arguably more critical – risk: the opportunity cost of doing nothing. As countries like China move rapidly to develop frontier AI capabilities, failing to build equivalent systems could erode decision advantage, strategic awareness, and global influence.
The report doesn’t promise AI will make strategic warning effortless. But it does make the case that failing to explore and invest in AI-driven capabilities now could leave national security communities flying blind in an increasingly turbulent world. There is no single model, platform, or vendor that will solve this. What’s needed the authors say is a phased, multidisciplinary, technically rigorous strategy – one that prioritises data infrastructure, values human-machine collaboration, and sets realistic expectations.