Stylometric Signatures of AI Translation: Machine Learning Classification of GPT-4 and Human Translations Based on Grammatical Style

This study presents a digital stylistic investigation into AI-generated translations through the lens of machine learning. Leveraging a sentence-aligned Chinese-English UN corpus of 100,000 segments, we explore whether GPT-4 translations exhibit stylistic regularities that differ meaningfully from those produced by human translators. We extract stylometric and grammatical features—including sentence length, clause density, function word distributions, passivization, and modality—and use them to train classification models (Random Forest, SVM, Logistic Regression, and BERT). Results show over 90% accuracy in distinguishing GPT-4 from human translations, with interpretability analyses (SHAP) highlighting key discriminative features. These findings contribute to the expanding field of AI-authorship attribution, demonstrating how stylistic signals can serve as computational fingerprints for large language model (LLM) outputs. We argue for a digitally grounded approach to translation style analysis—one that not only enhances AI translation detection but also reshapes how we understand authorship, voice, and textual agency in the age of generative language models.