Penerapan Model Pre-Trained BERT dalam Mendeteksi Teks Buatan ChatGPT

The current development of AI has both positive and negative impacts. One of the positive impacts of AI, which can generate text, is that it makes it easier for humans to summarize and create text. However, there are also several negative impacts, such as the potential for misinformation, possible bias in content creation, risks to user privacy, and the possibility of producing incorrect information that could be harmful. This research uses the BERT method to detect text generated by ChatGPT. The stages of this research include data collection, preprocessing, dataset splitting, BERT classifier, and model evaluation using a Confusion Matrix. The dataset is divided into two categories: texts with less than 200 words and texts with more than 200 words. The test results show that with the SGD optimizer, the dataset with more than 200 words achieved higher accuracy (97%) compared to the dataset with less than 200 words (95%). Meanwhile, when using the AdamW optimizer, both datasets achieved the same accuracy, 97%, but with a lower loss for texts with less than 200 words (0.86%) compared to texts with more than 200 words (18%).