Abstract
The rise in the development of giant language models has also led to the need for a reliable approach in distinguishing human-generated text from AI-generated text, particularly across various languages. However, the majority of the current approaches to text detection have largely remained limited to the English language, thereby being ineffective in languages that have a different structure or limited resources. The current research proposes a remedy to this problem by developing and evaluating a cross-lingual text detection model based on multilingual transformer models. For the purpose of fine-tuning the models, the team utilized a balanced dataset consisting of 98,224 samples in nine different languages to determine whether the text was human-generated or AI-generated. Out of the models, the performance of the XLM-RoBERTa model was the best, achieving a result of 95.56% accuracy and 96.29% F1-score after the application of the threshold optimization technique. The model also demonstrated excellent performance in all the languages. However, the accuracy in the English language remained slightly lower because of the diversity in the writing style.