This study investigates whether large language models (LLMs) recognize authorial style across English-to-Japanese translation and aims to distinguish memorization-based from abstraction-based explanations of model performance. In an early experiment conducted shortly after the release of ChatGPT, the model attributed a passage from Cormac McCarthy's The Road, presented in Japanese translation, to Haruki Murakami (Author 2023). Since then, frontier models have dramatically improved, yet freely accessible AI tools still produce confident misattributions. This persistent gap raises a broader question: does authorial style survive translation, and when models succeed, are they recognizing style or recalling memorized associations?

Focusing on nineteenth-century American narrative prose, the study tests both frontier commercial models and freely accessible AI tools on a corpus of six works by Irving, Poe, Hawthorne, Douglass, Melville, and Alcott. Each excerpt is prepared in three conditions: the English original (EN), a published Japanese translation (JP-HT), and a machine translation (JP-MT). Because the English originals are in the public domain and widely available online while their Japanese translations largely are not, the three conditions create a scale of likely data exposure.

The experimental design combines LLM-based attribution under controlled prompting, a masked condition replacing culturally specific markers with neutral placeholders, and a memorization test using n-gram overlap metrics. Multilingual Sentence-BERT provides a complementary embedding-based perspective. The study places particular emphasis on the qualitative analysis of model errors, examining misattribution patterns through the lens of literary history and reception.