Building Epistemic Coalitions: How Cross-Sector Alliances Between Academia, Industry, and Civil Society Shape AI Governance Norms

AI governance norms do not emerge through top-down regulation alone. Before formal legislation arrives - and often determining what legislation looks like when it does - governance norms are shaped by informal coalitions that establish shared language, produce evidence, define problem boundaries, and create policy templates that regulators subsequently adopt, adapt, or contest. This paper maps how these epistemic coalitions - cross-sector alliances between academic researchers, industry practitioners, civil society advocates, and policy entrepreneurs - actually function in the AI governance space. Drawing on epistemic community theory from international relations (Haas, 1992) and co-production frameworks from science and technology studies (Jasanoff, 2004), the paper introduces the Epistemic Coalition Architecture Model (ECAM) - a framework identifying four structural features that determine whether cross-sector coalitions produce governance norms that serve broad public interest or are captured by the most resource-rich participants. The model is validated through comparative analysis of four norm-formation episodes: the development of AI fairness metrics and benchmarks (academia-led, industry-adopted), the establishment of AI transparency norms (civil society-led, partially legislated), the construction of AI safety narratives (industry-led, contested by researchers), and the emergence of AI sovereignty discourse (government-led, industry-resisted). Analysis of 31 coalition-level interactions across these episodes reveals that coalition effectiveness depends on four structural features: shared evidentiary standards, credible boundary-spanning individuals, institutional platforms for sustained interaction, and governance of resource asymmetry between coalition members. The paper argues that epistemic coalitions are the invisible infrastructure of AI governance and that understanding their dynamics is essential for anyone seeking to shape - or resist - the norms that will govern algorithmic systems.