A machine learning model designed to predict loan approvals demonstrates high accuracy on internal test data but performs poorly when evaluated on data from new applicant demographics not well represented in the training set. What is the most likely cause of this issue?