Constructive journalism is a genre that aims to solve the problem of news avoidance, which can have serious social implications. One major reason is the negative focus of the news cycle. We investigate if machine learning models can perform constructive and non-constructive classification of news articles. The state-of-the-art BERT language model was compared with four traditional machine learning methods: logistic regression, random forest, gradient boosting, and support vector machine. The traditional models were trained on both metadata and TF-IDF. Lastly, explainable AI was implemented to gauge whether the models were trustworthy. The BERT model achieved an accuracy of 81.25%. On accuracy, it was outperformed by SVM on metadata (87.50%) and random forest on word embedding (89.58%). However, when using BERT we are able to make more useful explanations of the model. Due to the fact that it was able to consider the context of words, whereas the weights of features are constant in traditional methods.