Large Language Models are of fundamental importance for future AI usage in industry and society. The environmental impact of training and inference of such models however is devastating, which necessitates research into more efficient model mechanisms. In the simple context of the well-established BERT architecture, this study performs a comprehensive comparison between classical fine-tuning, which involves most or all of the model weights, and parameter-efficient Adapter Tuning, in which shallow trainable layers are introduced throughout the model while the vast majority of original weights remain unchanged. A series of experiments were carried out in which the BERTbase-architecture was trained on the LexGLUE benchmark with fine-tuning and Adapter Tuning for each data set. Through extensive comparisons in the experiments, it was found that Adapter Tuning is advantageous from a green AI perspective at epoch level, supporting the claims made in the literature. However, in our experiments reaching model convergence with Adapter Tuning requires significantly more time, making this training method in total less environmentally friendly than the conventional fine-tuning method, at least for the chosen model family.