There is a wide range of possible undesirable outcomes in ML. These can range from models that simply don't work to models that do work but use an unnecessary amount of resources in the process. Negative outcomes can be caused by many factors, such as the selection of an inappropriate algorithm, poor feature engineering, improper training techniques, insufficient preprocessing, or misinterpretation of results.
In the best-case scenario—that is, the best-case scenario of a negative outcome—the problem will make itself apparent in the early stages of your implementation. You may find during the training and validation stage that your ANN never achieves an accuracy greater than 50%. In some cases, an ANN will quickly stabilize at a value like 25% accuracy after only a few training epochs and never improve.
Problems that make themselves obvious...