Robust detection of AI-generated text using stylometric-semantic modeling under paraphrasing and adversarial rewriting

Abstract

The widespread usage of large language models like ChatGPT, Claude and Gemini has made it harder to distinguish between human and AI-generated writing. This research describes a fully reproducible and complete AI-generated text recognition pipeline that uses stylistic and semantic (embedding-based) properties to detect text well. The public dataset, Human vs. LLM Text Corpus was used for this research. This study can examine its proposed detection methods at a realistic sample size because this dataset has 25 times more samples than previous studies. The research analyzes adversarial robustness against paraphrase assaults and evaluates TF-IDF-based classifiers, sentence-embedding models and hybrid fusion architectures. The TF-IDF baseline outperformed embedding-based approaches with an accuracy of 83.14% and a ROC-AUC of 91.87% on the complete unbalanced dataset. Using a balanced dataset, a hybrid LightGBM model obtained 80.98% accuracy and showed strong robustness, with just a 2.52% loss in F1-score after paraphrase attacks. Document-level stylistic features dominated the model's decision-making process, with word count being the most discriminative variable (importance = 646), despite accounting for less than 1% of all features. This study presents a clear, large-scale benchmark for detecting AI-generated content and shows that feature interpretability and structural indications are needed for accurate AI authorship verification.