Due to their outstanding elastic limit, biocompatible Ti-based bulk metallic glasses (BMGs) are candidate materials to decrease the size of medical implants and therefore reduce their invasiveness. However, the practical use of classical Ti-BMGs in medical applications is in part hindered by their high copper content: more effort is thus required to design low-copper Ti-BMGs. In this work, in line with current rise in AI-driven tools, machine learning (ML) approaches, a neural-network ML model is used to explore the glass-forming ability (GFA) of unreported low-copper compositions within the biocompatible Ti-Zr-Cu-Pd system. Two types of models are trained and compared: one based on the alloy composition only, and a second based on various features derived from the alloying elements. Contrary to expectation, the predictive power of both models in evaluating GFA is similar. The compositional space identified by ML as promising is experimentally assessed, finding unfortunately low GFA. These results indicate that the ML approach may be premature for specific composition tuning of amorphous metallic materials. We emphasise that the development of ML tools in GFA prediction requires an improvement of the dataset, in terms of homogeneity, size and GFA descriptors, which must be supported by increased reporting of high-quality experimental GFA measurements, both positive and negative. Statement of significanceBiocompatible Ti-based bulk metallic glasses (BMGs) are candidate materials for use in the next generation of minimally invasive dental implants where improved mechanical properties, such as high strength are required. Despite promising in vitro/vivo evaluations, implementation of alloys for practical applications is partly hindered by the presence of copper as the main alloying element. Recent studies have presented AI-guided and machine learning strategies as appealing approaches to understand and describe the glass forming ability (GFA) of BMG-forming compositions. In this work, we employ and evaluate the capacity of a machine-learning model to explore low-copper compositional spaces in the biocompatible Ti-Zr-Cu-Pd system. Our results highlight the limits of such a computational approach and suggest improvements for future designing routes.