Generative AI Governance Models

Abstract

Generative AI -- foundation models capable of producing text, images, audio, video, and code at human-competitive quality -- has created governance challenges that existing regulatory frameworks were not designed to address. Traditional product safety regulation assumes identifiable manufacturers, deterministic product behaviour, and bounded application domains. Generative AI systems are general-purpose, non-deterministic, trained on vast datasets of uncertain provenance, deployed through complex value chains of developers, API platforms, and downstream integrators, and capable of producing outputs across virtually every domain of human activity. This study comparatively evaluates six generative AI governance models: the EU AI Act (risk-based, horizontal), the US Executive Order on AI (sector-based, voluntary-first), the UK AI Safety Institute model (technical evaluation focus), the China AI governance model (registration and content control), Singapore's Model AI Governance Framework (principles-based), and a proposed Adaptive Layered Governance Framework (ALGF) that integrates technical, legal, and market mechanisms. Evaluation criteria include comprehensiveness, legal enforceability, innovation impact, international compatibility, and implementation feasibility. Structured assessment by 48 legal, technical, and policy experts across six criteria reveals that no existing framework is comprehensive on all dimensions. The EU AI Act achieves the highest comprehensiveness (4.24/5) but the lowest innovation impact score (2.84/5). The US model achieves the highest innovation score (4.12/5) but lowest enforceability (2.42/5). ALGF achieves the best balanced profile (mean 3.84/5). A governance convergence roadmap for international generative AI regulation is proposed.