Identifying AI-Generated Text Designed to Evade Plagiarism Detection

It is inevitable to detect such content to ensure intellectual and academic integrity. Presented study investigates robust approach toward the detection of AI-generated content using a fine-tuned BERT model. A labeled Kaggle dataset with human- and AI-generated text has been used for training, validation and performance evaluation. Data preprocessing has been carried out using text normalization, noise removal, and tokenization using the BERT tokenizer. To address the memory constraint, the dataset has been processed in batches. The model has been fine-tuned with sparse categorical cross-entropy loss and the Adam optimizer. Classification in the form of SoftMax probabilities using a threshold to classify text either as human generated or AI-generated form the basis for performance evaluation. The presented model demonstrates the superior performance in comparison to Roberta and DistilBERT models for all the four evaluation criteria with accuracy of 72%, precision of 80%, recall of 60% and F1- score of 69%. This investigated scalable framework paves way for effective and efficient detection of plagiarism through AI-generated content and humanizers.