Knowledge, Innovation and Technology in Scientific Research Vol. 5

Vehicular ad hoc networks (VANETs) must sustain communication among rapidly moving vehicles and infrastructure despite short-lived links, heterogeneous radio access, fluctuating traffic density, stringent latency and reliability requirements, and persistent security threats. Artificial intelligence (AI) and machine learning (ML) are therefore being investigated as adaptive mechanisms for functions that conventional rule-based or optimisation approaches handle poorly under uncertainty. This critical narrative review evaluates how learning-enabled methods are being used in VANET and vehicle-to-everything (V2X) networking, with emphasis on routing and clustering, congestion and traffic prediction, radio-resource allocation, proactive caching, vehicular edge computing, federated learning, intrusion and misbehaviour detection, and emerging graph-based and multi-agent approaches. Literature published from 1 January 2000 to 1 August 2026 was considered, with conceptually important foundational evidence retained when necessary within this boundary. The evidence indicates that reinforcement learning and deep reinforcement learning are particularly well aligned with sequential resource, routing and offloading decisions, while supervised and ensemble methods remain useful for prediction and attack classification. Federated learning can reduce raw-data centralisation, and graph neural networks offer a promising representation of topology-dependent decisions. Nevertheless, most reported gains remain simulation-dependent and are difficult to compare because mobility, radio, attack, workload and baseline assumptions vary widely. Data drift, non-independent vehicular data, unstable client participation, adversarial manipulation, communication and computation overhead, limited interpretability, and weak cross-scenario validation constrain deployment claims. The review therefore argues that the central research problem is no longer whether ML can improve a selected VANET metric in a controlled experiment, but whether learning-enabled network functions remain safe, efficient and transferable under realistic mobility, mixed radio technologies, adversarial conditions and resource constraints. Progress requires common evaluation protocols, cross-domain validation, uncertainty-aware and safe learning, communication-efficient distributed intelligence, and closer integration of networking, security and edge-computing objectives.