The GPT Dilemma: Foundation Models and the Shadow of Dual-Use

TL;DR

This paper tackles the problem that foundation models blur civilian and military boundaries, threatening international stability. It introduces a four-factor framework—model inputs, capabilities, system use cases, and system deployment—to assess when civilian and military uses are distinguishable and how this distinguishability evolves during development. By applying the INF Treaty as a case study, it proposes concrete strategies: establish red lines for military competition, enhance information sharing, develop verification through national technical means, and constrain weapon platforms or deployments. The analysis emphasizes that dual-use risks arise at every stage, and improving distinguishability can enable safer competition and more credible security trades. The work aims to guide policymakers and AI labs toward responsible innovation that preserves strategic stability while realizing the economic and scientific benefits of general-purpose AI technologies.

Abstract

This paper examines the dual-use challenges of foundation models and the consequent risks they pose for international security. As artificial intelligence (AI) models are increasingly tested and deployed across both civilian and military sectors, distinguishing between these uses becomes more complex, potentially leading to misunderstandings and unintended escalations among states. The broad capabilities of foundation models lower the cost of repurposing civilian models for military uses, making it difficult to discern another state's intentions behind developing and deploying these models. As military capabilities are increasingly augmented by AI, this discernment is crucial in evaluating the extent to which a state poses a military threat. Consequently, the ability to distinguish between military and civilian applications of these models is key to averting potential military escalations. The paper analyzes this issue through four critical factors in the development cycle of foundation models: model inputs, capabilities, system use cases, and system deployment. This framework helps elucidate the points at which ambiguity between civilian and military applications may arise, leading to potential misperceptions. Using the Intermediate-Range Nuclear Forces (INF) Treaty as a case study, this paper proposes several strategies to mitigate the associated risks. These include establishing red lines for military competition, enhancing information-sharing protocols, employing foundation models to promote international transparency, and imposing constraints on specific weapon platforms. By managing dual-use risks effectively, these strategies aim to minimize potential escalations and address the trade-offs accompanying increasingly general AI models.