TLDR: This research introduces a trustworthy co-learning model for human-AI teaming in military operations, addressing the critical need for effective AI integration in defense systems. Developed using System Dynamics, the model features adjustable autonomy, multi-layered control, bidirectional feedback, and collaborative decision-making. A use case involving proportionality assessment in military targeting demonstrates that while the co-learning architecture can achieve a lawful balance between military advantage and collateral damage, this equilibrium is highly vulnerable to environmental uncertainty. The model effectively quantifies the trade-offs between AI delegation and legal robustness, providing a foundation for developing adaptable and ethically compliant human-AI partnerships.
In today’s rapidly changing world, military operations face increasingly complex threats. Artificial Intelligence (AI) offers significant advantages in areas like threat classification and decision support, but its integration also brings challenges related to effectiveness and ethics. This is where the concept of human-AI teaming becomes crucial, focusing on how humans and AI can work together as partners, combining human judgment and creativity with AI’s data processing and automation capabilities.
While human-AI teams are already used in various fields, from cancer detection to manufacturing, their application in military operations is particularly critical. Here, human operators and AI systems collaborate to analyze battlefield data and inform tactical decisions. A key aspect for effective human-AI teaming, especially in high-stakes military environments, is ‘co-learning’ – the continuous, mutual adaptation and learning between human and AI agents. This process helps build shared understanding, trust, and ensures operational effectiveness and safety in unpredictable situations.
Despite the growing interest, there has been a gap in scientific literature regarding models that explicitly capture and represent these co-learning processes. To address this, researchers Clara Maathuis and Kasper Cools propose a trustworthy co-learning model specifically designed for human-AI teaming in military operations. Their model aims to foster a continuous and bidirectional exchange of insights, allowing both human and AI agents to adapt jointly to evolving battlefield conditions.
The proposed model integrates four key dimensions to achieve this:
Adjustable Autonomy
This dimension allows for the dynamic calibration of AI agents’ autonomy levels. Depending on factors like the mission state, the AI system’s confidence, and environmental uncertainty, the AI’s independence can be adjusted. This ensures flexibility and context-sensitive teamwork.
Multi-layered Control
This involves continuous oversight and monitoring of activities by both human and AI agents. It promotes transparency, accountability, and helps ensure operational safety by allowing intervention at various levels.
Bidirectional Feedback
The model emphasizes explicit and implicit feedback loops between agents. This ensures proper communication of reasoning, uncertainties, and learned adaptations from both the human and the AI, fostering mutual understanding.
Also Read:
- AI Combat Pilots: Enhancing Human-Agent Teaming in Air Combat Simulations
- Unpacking the Architectural Shift: AI’s Impact on Cyber-Physical Systems and Verification
Collaborative Decision-Making
This dimension focuses on the joint generation, evaluation, and proposal of decisions. It includes confidence levels and the rationale behind decisions, enhancing transparency and preventing over-reliance or under-reliance on AI recommendations.
The model uses a System Dynamics (SD) approach, which is well-suited for understanding complex systems with interdependencies and feedback loops. It defines ‘stocks’ representing static characteristics like human expertise, AI competence, shared situation awareness, trust calibration, AI authority level, and cognitive load. ‘Flows’ represent the dynamic changes in these stocks, such as human learning, AI learning, and authority adjustment, driven by interactions and information exchange.
To demonstrate its practical applicability, the model was tested using a use case focused on the proportionality assessment in military targeting. This assessment involves weighing the expected collateral damage to civilians against the anticipated military advantage of an operation. In a hypothetical scenario involving a cyber-kinetic strike on an adversary’s air-defense network in an urban area, a human operator and an AI decision-support agent jointly performed this assessment.
The simulation revealed that while the co-learning architecture could initially achieve a lawful balance between military advantage and collateral damage risk, this equilibrium proved fragile under high environmental uncertainty. During the first 30 minutes, bidirectional feedback improved both human expertise and AI competence, aligning their mental models and calibrating trust. This led to a cautious increase in delegated authority and a manageable cognitive load for the human. However, when environmental volatility was introduced (representing dense civilian presence and potential disruption), trust collapsed faster than the oversight mechanisms could adjust AI autonomy. This triggered a rapid withdrawal of authority, causing the military advantage to fall sharply and the proportionality score to become negative for a significant portion of the mission window, indicating a failure to meet legal sufficiency criteria.
The results underscore the model’s effectiveness in reproducing expected doctrinal behaviors, such as autonomy retraction under uncertainty and workload-driven safety loops. It also quantifies the trade-offs between aggressive delegation of tasks to AI and maintaining legal robustness in military decision-making. This research provides a foundational framework for developing human-AI teaming systems that truly partner with human operators, adapting and learning together while upholding ethical standards and accountability in an era of evolving threats. For more details, you can read the full research paper here.
