Detection of Bangla AI-Generated Texts Using a Quantum Kernel Support Vector Machine

Abstract

Machine-generated text is increasingly hard to distinguish from human writing, raising concerns for academic integrity, journalism and information trust, especially for low-resource languages such as Bangla, where detection tooling lags far behind English. This paper presents, to the best of our knowledge, the first application of a quantum-kernel Support Vector Machine (QSVM) to detecting AI-generated Bangla text. On a class-balanced Bangla dataset spanning six genres (newspaper articles, scientific abstracts, emails, poems, short stories and office orders) that is strongly imbalanced across genres, we extract BanglaBERT contextual embeddings, reduce them to eight principal components and map each sample onto an 8-qubit Hilbert space with an angle-embedding feature map and a ring of entangling CZ gates. The fidelity between mapped states forms a quantum kernel that is passed to a classical SVM. On a held-out test split, the QSVM attains 95.0% accuracy, 95.0% F1-score and 0.989 ROC-AUC. Under the same 5-fold stratified cross-validation protocol applied to all models, it achieves 95.2% (±2.8%) mean accuracy, compared with 95.0% (±2.0%) for Logistic Regression and 93.8% (±2.4%) for an RBF-kernel SVM on identical features and 85.8% (±4.1%) for a Variational Quantum Classifier (VQC). Replacing BanglaBERT with TF-IDF features lowers the QSVM's cross-validated accuracy to 93.7% (±2.7%). Within the specific configurations evaluated, the QSVM was more accurate and less variable than the single VQC architecture tested; the experiments do not identify the cause of this gap. Per-genre results are reported alongside the aggregate metrics. These results provide a first quantitative benchmark for quantum-enhanced Bangla AI-text detection and indicate that the QSVM is competitive with, rather than superior to, strong classical baselines.