How can big data and AI prevent violent conflict?

UCL Policy Lab and Chatham House's International Security Programme continue their partnership through the Innovation Network in Conflict Prediction with a second roundtable which brought together experts and policymakers. Held under the Chatham House Rule, this session focused on examining conflict prediction models, exploring their current capabilities, future potential, and ways to mitigate algorithmic biases.

The meeting also considered critical questions about the relationship between technology partners and the conflict contexts their systems might operate in, particularly when technical expertise may not be matched with deep contextual understanding.

The session revealed important insights about the methodology of collecting data and goals of using conflict prediction models:

• Predicting how existing conflicts will intensify and change is often more valuable and accurate than attempting to predict entirely new conflicts. For policymakers, understanding which dormant conflicts might escalate is a key priority, with a focus on predicting potential escalation at least one year in advance.
• More innovation is needed in thinking about which types of data predict political instability. While currently undervalued and consequently unincorporated, local, “on-the-ground” data is key to improve conflict prediction.
• Researchers and policymakers need to be aware of biases to avoid recreating problematic dynamics from the past, particularly around data ownership and interpretation.
• Developing and maintaining good transparency practices is essential for achieving maximum data coverage and consistency.
• Using multiple, potentially biased sources can create a more accurate picture when properly cross-referenced, though this requires resource-intensive local expertise to mitigate the biases.
• It remains questionable whether predicting events and fatality counts is the most useful approach, or if predicting actor behaviour and political instability offers more valuable insights.

Participants also identified fundamental challenges in data collection and conflict prediction:

• Resource constraints and the dismantling of USAID may result in less data collection, creating a need to "do more with less."
• Conflict data often lacks adequate local reporting (partially due to insufficient resources) and may not be accessible to the communities experiencing the conflict.
• Balancing the need for specific, contextual information with consistency across different conflict situations remains difficult.
• The sector lacks clear guidelines for what data should be shared and how, hampering cooperation.
• Growing awareness of data misuse may lead to decreased willingness among populations to share information, requiring more trust-building measures as well as transparency about privacy protections and data protection measures.
• While well-designed models are valuable, ensuring policymakers take action based on predictions remains challenging.
• Current models excel at predicting incidents of conflict but struggle to predict new conflicts or significant changes in existing ones.

The second meeting of the Innovation Network reinforced the importance of the research-policy interface in addressing complex conflict situations. The discussions highlighted the need for balance: between precision and accessibility, between technical capability and contextual understanding, and between competing priorities in the conflict prediction space. As the global conflict landscape continues to evolve, collaborative approaches that bridge these divides will be essential for developing truly effective early warning systems. The workshop produced several recommendations for improved conflict prediction models and effective utilisation of data:

• Organisations must revisit assumptions about what their models and data collection are intending to achieve, with greater clarity about specific aims. Conflict prevention is not usually an organisation’s mandate, and this should be made clear.
• Rather than competing for the "best" system, different or complementary forecasting approaches may prove most effective.
• Policymakers need concise, actionable information - typically one-page summaries that are straight to the point.
• There is a need for diagnostic indicators at earlier stages, even if they come with lower confidence levels.
• Different stakeholders have different needs from prediction systems, requiring customised outputs prioritising precision vs recall.
• Providing information on how and why predictions change within models can build trust with decision-makers.

 This work was supported by UCL Innovation & Enterprise.

Author: Georgia Cole (Research Analyst, International Security Programme), with contributions from Manuel Vogt (Associate Professor, Department of Political Science).