Detection of AI-Generated Text Using Machine Learning and Deep Learning

Abstract:

The growing fluency of large language models has made AI-generated text increasingly difficult to distinguish from human writing, raising direct consequences for academic integrity, editorial review, and platform trust. Existing detection approaches fall into two largely separate families: statistical and stylometric methods that exploit token-level probability and burstiness signals, and supervised machine learning and deep learning classifiers, typically fine-tuned transformer encoders, trained directly on labeled corpora. This paper proposes a hybrid detection pipeline that fuses statistical/stylometric features with fine-tuned transformer embeddings through a calibrated ensemble, followed by an explicit adversarial paraphrase-robustness check before deployment. Drawing on established AI-text-detection and adversarial-robustness literature, the study defines seven evaluation criteria — Detection Accuracy, Precision, Recall, F1 Score, False Positive Rate, AUC-ROC, and a combined Detection Reliability Score — and applies them to an illustrative case study of an academic-integrity screening deployment. A five-configuration ablation study, spanning a classical bag-of-words baseline through a deliberately over-engineered ensemble, is used to identify the point of diminishing returns, and the recommended configuration is benchmarked directly against the classical baseline, including accuracy under a simple paraphrasing attack. The illustrative results indicate substantially higher accuracy and F1 score, a markedly lower false-positive rate, and materially better retained accuracy under paraphrasing for the proposed hybrid pipeline relative to the baseline. The findings position statistical-plus-learned ensemble detection, evaluated against explicit adversarial criteria, as a decision-support tool for human review rather than an infallible automated verdict, and as a foundational deployment practice for institutions and platforms screening for AI-generated text.