WHEN FAILURES TRAVEL: HOW NEGATIVE RESULTS IN OPEN AI RESEARCH ENABLE MILITARIZATION

Research on AI militarization has largely focused on the transfer and deployment of successful models and systems. In contrast, the role of negative experimental outcomes—including failed models, abandoned datasets, and documented limitations—has received little systematic attention. In this paper, we present a preliminary sociotechnical analysis of how failure disclosures in open AI research circulate beyond their original academic contexts and become informational resources within military and security research ecosystems. We focus exclusively on knowledge flows and governance implications, rather than evaluating military AI systems or proposing technical improvements. We argue that prevailing norms of transparency around failure reporting, while essential for scientific progress, can unintentionally lower barriers for militarized reuse and create informational dual-use risks similar to those observed in other scientific fields. We conclude by outlining governance-oriented considerations for handling failure disclosures in high-risk AI research domains and suggest avenues for fostering more reflexive research practices that acknowledge these dual-use dilemmas.