Study reveals LLMs are erasing linguistic fingerprints of non-native English speakers in academic writing, though some languages resist homogenization

Researchers analyzed native language identification trends across ACL Anthology papers spanning three eras: pre-neural network, pre-LLM, and post-LLM periods

Native language identification performance consistently declined over time, suggesting writing assistance tools are smoothing out author linguistic backgrounds

Post-LLM era shows unexpected anomalies: Chinese and French papers maintained or diverged from expected trends, while Japanese and Korean showed steeper-than-expected declines

Study used a semi-automated framework and fine-tuned classifier to detect linguistic fingerprints revealing how authorial native language signals persist or disappear in academic writing