AI Erases Linguistic Fingerprints in Academic Writing
A Quiet Erasure in Academic English
Somewhere in the machinery of academic writing, a subtle transformation is underway. Researchers submit their abstracts to journals and conferences, and increasingly, those abstracts pass through large language models that smooth, polish, and normalize. The result reads cleanly. It reads professionally. And according to new research, it reads less and less like the person who wrote it.
A study examining native language identification — the ability to detect an author’s first language from their writing — has found that the signals distinguishing, say, a German researcher’s English from a Japanese researcher’s English have been steadily fading. The finding matters not because diversity is a nice thing to have, but because it reveals something uncomfortable about what these systems actually do versus what we believe they do. We think they help us communicate. The evidence suggests they also flatten us into something more uniform than we intended.
What the Data Shows
The research, published in the ACL Anthology, constructed two datasets of academic abstracts drawn from arXiv and the ACL Anthology, covering eight native language groups across three distinct periods: before neural networks became dominant, before large language models arrived, and after their widespread adoption The researchers then trained classifiers to identify authors’ native languages and measured how well those classifiers performed in each era.
The results show a consistent decline over time. But here is where the story becomes more interesting than a simple narrative of “AI ruined everything.” The drop in identifiable native language signals was actually steeper between the pre-neural-network era and the pre-LLM era than between the pre-LLM era and the post-LLM period. This means the homogenization did not begin when ChatGPT appeared. It has been progressing gradually since neural approaches to language processing first took hold.
The deception here is not that LLMs are secretly erasing our linguistic identities. It is that we have been telling ourselves a story about when and why this happened — and the timeline does not match the evidence.
The Rewriting Experiment
To test how much further this could go, the researchers conducted a rewriting experiment using recent LLMs. They took existing texts and had the models revise them, then measured how much native language signal remained. The drop was larger than the entire progression across historical eras combined.
This is the gap between appearance and reality in its clearest form. A researcher uses an LLM to “improve” their abstract. The text becomes more fluent, more idiomatic, more aligned with what the model considers standard academic English. From the researcher’s perspective, the tool has done exactly what it promised: made the writing better. But “better” in this context means more like everyone else. The model is not enhancing the author’s voice. It is replacing it with a statistically average voice that belongs to no one.
The tool does not announce this. It presents its output as an improvement, and in the narrow sense of readability, it is. What gets lost is invisible unless you know to look for it.
Why This Matters Beyond Linguistics
The implications extend past academic writing. Every domain where LLMs assist human expression — journalism, legal writing, corporate communication, personal correspondence — faces the same dynamic. The model optimizes for what its training data suggests is appropriate. That training data reflects the dominant conventions of the language. The result is text that passes smoothly through whatever gatekeeping mechanism it encounters, whether that is a journal editor, a hiring manager, or an algorithm that ranks search results.
What the model cannot preserve is the particular way a non-native speaker structures an argument, the rhythm that comes from thinking in one language and writing in another, the small idiosyncrasies that signal a human being with a specific history. These are not flaws to be corrected. They are information.
The research suggests that native language identification performance has declined so much that in some cases, the classifiers struggle to distinguish between authors from different linguistic backgrounds at all. The fingerprints are still there, but they are fainter. And with each pass through an LLM, they grow fainter still.
The Question We Are Not Asking
The study raises a question that goes beyond whether we can still trace L1 signals. It asks whether we should want to. The implicit assumption in much of the discussion around AI writing assistance is that homogenization is an acceptable price for clarity and accessibility. Everyone writes in the same standardized English, and communication becomes easier. The friction of encountering unfamiliar sentence structures disappears.
But friction is not always an obstacle. Sometimes it is a sign that a real person is on the other end, someone who sees the world through a different linguistic lens. The efficiency we gain by smoothing that away comes at a cost we have not fully calculated.
The researchers do not offer a solution, and perhaps there is not one that preserves both the convenience of LLM assistance and the distinctiveness of individual voices. What they offer instead is a measurement — evidence that something is being lost, and that the loss is accelerating. Whether that matters is a question we have to answer for ourselves, ideally before the signals are gone entirely.
The tools will keep improving, making our writing cleaner, more consistent, more aligned with whatever standard they have learned. The question is whether we notice what we are trading away — and whether we decide it is worth the price.