Human uses social media platform such as Twitter to express feelings and opinions through text about the surrounding issues. Understanding emotions at the subtle level of expressed feelings are essential for better human and computer interactions. The previous emotion recognition approach required many training data and lexical databases. Unfortunately, the availability of very little labeled training data is a limitation and challenge to achieving high model performance. Therefore, we investigate the BERT language model for emotion recognition in Indonesian-language Tweets in this study. We choose to use fine-tuning instead of pre-training, which requires extensive data and resources. Two pre-trained models were used to determine the effectiveness and performance of the proposed model. Experiments show that the proposed model outperforms all existing baseline models, with the highest accuracy is 77%. Another advantage that we analyze is that BERT requires a relatively short computation time. In addition, BERT has a better context representation.