A Survey of Large Language Model Use and Its Technical Limitations in Military Systems Through a Decolonial Lens (Extended Abstract)

Abstract

The increased integration of LLMs into military systems requires a decolonial analysis to evaluate the lasting sociopolitical impact of such systems on individuals in global conflict settings. In this paper, we first survey all publicly-identifiable use cases of LLMs across all military domains, then apply Edward Said's framework of Orientalism as a basis for our decolonial framework to analyze the longstanding technical limitations of LLMs within these domains. Through technical limitations such as bias, accountability, explainability, sycophancy, data scarcity and data colonialism, deficiencies in real-world understanding and commonsense reasoning, Western-centric ethics, and susceptibility to hallucination and misinformation, and through Saidian concepts of 'Othering,' the power of representation, and the silencing of marginalized perspectives through regulation of autonomy, the paper finds that the use of LLMs in the military serves to automate and amplify historical colonial dynamics of power, control, and purposeful misrepresentation of non-Western cultures through the weaponization of language. This analysis highlights the urgent need to address how LLM vulnerabilities in the military severely harm populations where such technology is used against, reproducing global systemic inequalities and colonial legacies within modern warfare.