Academic English has been quietly losing its accents for two decades, and LLMs are just the latest chapter.
Researchers built two datasets of academic abstracts from arXiv and the ACL Anthology, covering eight native-language groups across three eras: before neural networks, before LLMs, and after LLMs. They fine-tuned LLMs to detect native-language identification (NLI) signals - the subtle grammatical fingerprints that reveal a writer's first language - in each period. NLI accuracy dropped steadily over time, but the sharpest decline happened between the pre-neural-network and pre-LLM eras, not after LLMs showed up. A separate test where the team had recent LLMs rewrite the same abstracts erased more of those signals than the entire historical decline across all three eras combined.
The finding complicates the tidy story that chatbots are singlehandedly flattening how non-native English speakers write. Homogenization was already underway before generative AI existed, likely as spell-checkers, grammar tools, and earlier neural language models nudged writers toward the same conventions. Still, the rewriting result is the number worth watching: it suggests that as LLM-assisted writing becomes routine, the flattening could speed up considerably from here.
Call it the ghostwriting effect - not new, just newly automated and available to anyone with a deadline.