INFOPLAY: Local Generative AI for Adaptive Information Environment Simulation in Crisis Management and Civil Preparedness Exercises
Modern crisis management and civil preparedness exercises increasingly require participants to operate in complex and rapidly changing information environments. However, information injects are commonly pre-scripted, static, and labour-intensive to prepare, limiting the ability of exercise directors to adapt scenarios in response to participant actions or reproduce the volume, tempo, and competing narratives associated with contemporary hybrid threats.
INFOPLAY is a browser-based, portable simulation platform designed to support the controlled generation, management, and delivery of realistic information injects. The platform combines configurable actor profiles, geographic visualization, live situation feeds, scheduling, and locally hosted generative AI. Exercise staff can create and broadcast news articles, wire bulletins, social media posts, and authority communications from different actor perspectives, including public-service media, police, local authorities, bloggers, activist networks, and deliberately misleading information actors. Each actor profile defines communication style, objectives, target audience, channels, and credibility, enabling the same underlying event to be represented differently across the exercise information environment.
A key design principle is local deployment. INFOPLAY can use locally hosted large language models through platforms such as Ollama and LM Studio and can operate without cloud services or an external network connection. Generated content remains subject to human review, editing, scheduling, and approval. If local AI is unavailable, curated templates and pre-prepared content provide a fallback mechanism, supporting operation in restricted or air-gapped training environments.
An initial LiveX exercise deployment was used to assess the workflow in practice and compare AI-assisted production with the existing manual process. The platform was operated locally to manage scenario events, generate and schedule content, and present outputs through operator, participant, and geographic views. Observations indicated that generative AI can substantially reduce the effort required to transform a higher-level scenario narrative into multiple related exercise events and injects. At the same time, the trial confirmed that human validation remains essential for maintaining operational, pedagogical, and narrative coherence. Reliable offline packaging of the application and its dependencies was also identified as an important practical requirement.
These findings have shifted development from isolated content generation toward AI-assisted playbook generation, in which higher-level events can be decomposed into chronological sub-events and checked for temporal dependencies, narrative consistency, and exercise intent before human approval.
The current civil-preparedness configuration includes 11 actor profiles, 13 exercise scenarios across four escalation levels, and more than 110 curated injects, providing a structured baseline for adaptive AI-supported exercise control. INFOPLAY demonstrates a practical approach to integrating locally deployable generative AI into crisis-management simulation while retaining human authority, auditability, and offline resilience. Future work will compare manual and AI-assisted playbook production in terms of workload, quality, consistency, and usability.