Detection of AI-Generated Deepfake Text in Academic Submissions Using Fine-Tuned RoBERTa: A Transformer-Based Binary Classification Approach

Abstract

The rapid proliferation of artificial intelligence (AI)–assisted writing tools in academic 
environments has introduced a critical challenge for higher education institutions, particularly 
in safeguarding academic integrity and authenticity of scholarly outputs. This study proposes 
a robust binary text classification framework for discriminating between AI-generated 
deepfake texts and human-authored academic documents, leveraging a fine-tuned RoBERTa 
architecture. The proposed model was systematically optimized through transfer learning and 
task-specific fine-tuning, resulting in substantial performance improvements over the 
baseline pre-trained RoBERTa model. Specifically, the fine-tuned model achieved a validation 
accuracy of 99.0% and a weighted F1-score of 0.990 as early as Epoch 2 during GPU
accelerated training. In contrast, the base RoBERTa model, without domain adaptation, 
yielded significantly lower performance, with an accuracy of 74% and an F1-score of 0.73. 
Additionally, training dynamics reveal a pronounced reduction in loss values, decreasing from 
0.2005 in early iterations to 0.0032 by Epoch 4, indicating rapid convergence and effective 
feature learning. The study presents a comprehensive experimental pipeline encompassing 
architectural design, dataset preprocessing, fine-tuning configurations, and model training 
protocols. Evaluation metrics include confusion matrix analysis, loss curve diagnostics, and 
comparative benchmarking against the baseline model. The proposed classifier demonstrates 
strong generalization capability on the RAID benchmark dataset, achieving balanced 
classification performance across both AI-generated and human-written classes, with only 17 
misclassifications out of 1,000 samples. Beyond performance, the paper also addresses ethical 
considerations surrounding AI detection systems, including implications for fairness, misuse, 
and policy integration. Overall, the results establish the proposed approach as a highly 
accurate and reliable solution for automated detection of AI-generated academic content.