Lexical and Rhythmic Fingerprints of Japanese LLM-Generated Text: A Cross-Model, Two-Layer Analysis

Which feature family carries the human/AI signal in Japanese LLM-generated text: word choice or rhythm? Practitioner tools claim that the "AI smell" of Japanese text resides in rhythm (monotone sentence-length variation) rather than vocabulary, but this claim has not been tested against lexical features under a controlled design. We compare lexical and rhythmic fingerprints on a corpus of 350 texts from 7 LLMs (Claude 3 Haiku, Claude Sonnet 4, Claude Opus 4, GPT-3.5 Turbo, GPT-4o, GPT-OSS 20B, Llama 3.2 1B) against 700 human technical-blog articles. We first rebuild the excess-vocabulary statistics of our prior study on a document-occurrence basis (Laplace smoothing, frequency ratio r and gap delta, document-level chi-squared tests with BH-FDR correction, bootstrap confidence intervals), following Kobak et al. (2025); this re-analysis is also a correction, as 35% of the 651 words we previously reported as significant do not survive at the document level. We then measure 13 interpretable rhythm metrics, including burstiness of sentence-length sequences and a mora-based approximation of phonological rhythm. Three findings emerge. (1) Rhythmic monotonization is real and shared: AI text shows lower burstiness (Cohen's d = -0.96), lower paragraph-structure variation (d = -0.67), and lower sentence-length variation (d = -0.56), with the same direction in all 7 models. (2) Contrary to the practitioner claim, lexical features dominate human/AI discrimination: cross-validated AUC is 0.998 for lexical vs. 0.897 for rhythm features, and the gap persists under document-length residualization and outlier-model exclusion. (3) At the multivariate level, the two families form a two-layer fingerprint: lexical profiles identify which model wrote a text (92% accuracy in 7-way model identification vs. 51% for rhythm), whereas rhythm features shift in a direction common to all models, with model-dependent magnitude. Rhythm forms the shared accent of machine writing; vocabulary is the signature of the individual model.