Abstract: Social media platforms like Twitter have become substantial sources of user-generated content, enabling people to easily express their emotions and opinions. However, this freedom has increased the spread of harmful content, such as abusive language, sexually explicit content, and hate speech. This poses challenges for content moderation and user safety. In order to guarantee a safer, more receptive, and more pleasurable online environment for users of all ages, it is essential to develop a system capable of recognizing abusive and sexually explicit material on Twitter. Despite the growing importance of content moderation, a research gap exists in Indonesian tweets, with limited comprehensive studies on negative content identification. This research addresses this gap by evaluating the effectiveness of Bidirectional Encoder Representations from Transformers (BERT) models in the Indonesian context, which were primarily developed for English and other languages. This research aims to identify abusive, adult, and neutral content in Indonesian tweets by examining and fine-tuning BERT-based models to maintain a healthy online environment for optimal tweet classification. Based on our experiments, the BERT-based models showed promising results in detecting negative tweets. Among the BERT-based models, IndoBERTweet achieved the best precision, recall, and macro F1 scores with 97.03, 96.88, and 96.94, respectively.