A new linguistics study finds that AI writing tools are systematically erasing the native language fingerprints of researchers in academic papers, with the post-LLM era accelerating a trend that began long before ChatGPT.
The paper, posted to ArXiv CS.CL, examines native language identification (NLI) — the ability to infer a writer's first language from patterns in their written English. Using papers from the ACL Anthology, a major archive of computational linguistics research, the authors built a labeled dataset and fine-tuned a classifier to detect these linguistic signals across three distinct periods: pre-neural network, pre-LLM, and post-LLM.
The post-LLM era reveals anomalies: while Chinese and French show unexpected resistance or divergent trends, Japanese and Korean exhibit sharper-than-expected declines.
How Researchers Detected a Disappearing Accent
The team constructed their dataset using a semi-automated labeling framework, then trained a classifier to identify linguistic patterns associated with specific native language backgrounds. The central metric is straightforward: how accurately can a model still identify a writer's origin based only on their prose? According to the study, that accuracy has declined consistently over time — and the drop steepens in the post-LLM era.
This approach treats writing like an accent. Just as spoken accents carry traces of a speaker's first language, written English from non-native speakers carries subtle syntactic and stylistic patterns. Grammar preferences, sentence construction habits, and idiomatic choices all leave traces that NLI classifiers can detect. The question the study poses is whether LLMs — trained overwhelmingly on standardised, predominantly native-English text — are smoothing those traces away.
According to the researchers, the answer is largely yes. But not uniformly.
Why Japanese and Korean Writers Are Most Affected
The study's most striking finding is the language-specific divergence in how this homogenisation plays out. Authors whose native languages are Japanese and Korean show sharper-than-expected declines in detectable NLI signal in the post-LLM era. This suggests that writers from these language backgrounds may be adopting AI assistance more heavily, or that the structural distance between their native languages and English makes LLM-polished output look more uniformly "standard."
Chinese and French native speakers, by contrast, show unexpected resistance — their linguistic fingerprints remain more detectable despite the same technological shifts. The study does not offer a definitive explanation for this divergence, but it raises important questions about differential adoption of writing tools and the structural properties of different language pairs.
This is a significant finding. If certain linguistic communities are more aggressively smoothing their written voice through AI tools, it could reflect economic or professional pressures — the perception that "native-sounding" English is a prerequisite for publication success in top venues.
The Broader Question of Academic Homogenisation
Beyond the technical linguistics, the study taps into a growing concern in academia: whether AI writing assistance is flattening the diversity of scholarly voice. Research writing is not purely functional. It carries intellectual tradition, rhetorical style, and cultural perspective. A French researcher and a Japanese researcher do not simply translate the same ideas into English — historically, their writing has reflected different modes of argumentation and exposition.
If LLMs are standardising that output toward a single, dominant register of academic English, the implications extend beyond linguistics. Peer reviewers, editors, and readers may increasingly encounter papers that sound alike regardless of their geographic or intellectual origin. Some researchers argue this is a levelling force — reducing disadvantage for non-native English speakers navigating a publication system that has historically penalised them. Others view it as a form of cultural erasure.
The study does not take a position on whether homogenisation is beneficial or harmful. Its contribution is empirical: it demonstrates that the signal is weakening, it quantifies the trend across eras, and it identifies which language communities are most affected.
Limitations Worth Noting
The dataset relies on a semi-automated labeling framework, which introduces potential noise in ground-truth native language attribution — a notoriously difficult task. The ACL Anthology is also a domain-specific corpus focused on computational linguistics and NLP, meaning findings may not generalise to other academic disciplines or non-technical writing. The post-LLM era in the dataset also represents a relatively short window, and the trend lines may shift as LLM adoption patterns evolve.
All benchmarks and findings referenced here are self-reported by the study authors and have not yet undergone formal peer review, as the paper is a preprint.
What This Means
For researchers, publishers, and anyone tracking AI's cultural footprint: the linguistic diversity of academic writing is measurably declining, and large language models are accelerating that process — unevenly, and in ways that differ significantly depending on a writer's native language.