Upper limb (UL) orthoses have long been used in rehabilitation to assist with movement and improve functional outcomes. However, current orthoses often lack the ability to provide individualized and adaptive assistance to each user. Machine learning (ML)-based orthoses may address these limitations by classifying movement patterns using biomedical signals. This review aimed to explore ML-based UL orthoses by describing the features, algorithms, and benefits from the data collected in articles and patents. We conducted a scoping review using multiple databases (PubMed, Web of Science, SciELO, Koreamed, AMED, CENTRAL, PEDro, IEEE Xplore, Scopus, and Arxiv), and patents websites (Patentscope, Patentlens, Google Patents Kripis, and J-platpat), and reference lists were reviewed to identify additional studies. Following the screening, the authors conducted data extraction (algorithm used, physical description of the orthosis, control system characteristics, and usability test performed); and the risk of bias (RoB) assessment, through prediction model study risk of bias assessment tool (PROBAST). Our review included 16 articles and 4 patents that met our inclusion criteria. Orthoses controlled by ML tools presented different features in the distinct aspects considered in this review. Regression models and surface electromyography signals were the most common approach in articles and patents. Unclear and high RoB data were identified in studies assessed through PROBAST. Our findings suggest ML-controlled orthoses provide benefits to UL movement. Nevertheless, larger and methodologically sounder studies should be conducted to assess the benefits of these devices.