Code smells signify suboptimal software design and implementation practices that can severely impact code maintainability. While traditional approaches to code smell detection have largely relied on heuristic and metric-based evaluations, recent advancements have explored the efficacy of Machine Learning (ML) techniques, specifically through the lens of Graph Neural Networks (GNNs) and Abstract Syntax Trees (ASTs). This paper critiques and synthesizes findings from two recent studies that employ these technologies to improve the automated detection of code smells. By tacking a close look to these approaches, we aim to highlight their contributions as well as their limitations within the context of current ML methodologies in software engineering. We provide a comparative analysis of the AST representations and GNN models utilized, exploring how they address the challenges of code smell detection and suggesting directions for future research. Our goal is to check the potential of these models to set new benchmarks in the field.