It is undeniable that artificial intelligence (AI) and machine learning (ML) have become rooted at every level of our society. It is also true that engineers have always strived for improvement in design, materials, and manufacturing, the three cornerstones encapsulating most engineering challenges. Hence, it is no surprise that recent years saw a surge in engineering contributions employing AI and ML techniques. However, conversely to analytical models and finite element or finite volume analyses (FEA, FVA), and despite the countless pros, AI and ML models also present several cons. Trying to avoid a lengthy analysis of all discernable aspects, this perspective focuses on two specific prospects: one inward and one outward-oriented issue, each representing a weak point for ML approaches but also a challenge. On the one hand, ML models’ formulations are well-known and documented. However, to achieve reasonable accuracy, the quality and size of the training dataset is paramount, making its definition as important as the architecture of the ML model itself. On the other hand, the prediction accuracy outside of the latent space, and how to improve it, remains a huge question mark, and often a limitation to more traditional yet reliable approaches, such as analytical, FEA, and FVA modeling. At this point, two questions arise. The former: what are the criteria to define whether ML has an edge over conventional modeling approaches? While the latter: how to design ML models capable of being less of a black box and better performing outside of the latent space? Each question is addressed separately in a section of the paper, together with a summary of the available state-of-the-art and a commentary of the authors’ perspective on the matter.