The primary use of supervised learning in machine learning is to train models using labeled datasets. Here, the algorithm learns from input-output pairs, which helps it make predictions or classifications on new, unseen data. This approach is commonly applied in areas such as image recognition and spam detection.