AI for NextGen 6G Networks: A Comprehensive Survey and Taxonomy

The transition from 5G to 6G networks brings unprecedented demands for massive connectivity, ultra-low latency, and exponentially higher data rates, rendering traditional, rigid Radio Access Networks (RAN) inadequate. To manage this exponential complexity, the integration of AI into the RAN—coupled with the disaggregated, open architecture of O-RAN—has emerged as a transformative paradigm, shifting network infrastructure from purely communication-centric entities to converged compute-communication platforms. This paper presents a comprehensive survey on the integration of AI within the 6G O-RAN ecosystem, focusing on the AI-RAN manifestations (AI-for-RAN, AI-on-RAN, and AI-and-RAN) and the end-to-end Machine Learning (ML) workflows driven by hierarchical RAN Intelligent Controllers (RICs). We propose a novel, multi-dimensional taxonomy that categorizes AI-RAN techniques by network function, AI paradigm, target use-case, and deployment strategy. Through a detailed analysis of state-of-the-art implementations, we highlight key insights, such as the efficacy of online Deep Reinforcement Learning (DRL) and GPU-accelerated edge computing, while identifying critical gaps, particularly the “black-box” nature of Deep Learning models that severely hinders operator trust. To address this, we emphasize the pioneering role of eXplainable AI (XAI) in achieving transparent and secure network automation. Finally, we outline open challenges such as multi-vendor interoperability and propose concrete future research directions, including Foundation Models at the telecom edge, Network Digital Twins, and Quantum ML, to realize fully autonomous, trustworthy 6G networks.