An analysis says the AI-specific writing style is seeping into human-written and edited text, which could reduce the reliability of methods that distinguish AI writing by words alone.

On Sept. 1 (local time), IT outlet TechRadar reported that computational linguist Fırat Mıhcı (피라트 미흐츠), founder of HumanizeMy.ai, analysed 18,989 abstracts posted on arXiv from 2019 to 2026 in computational linguistics, neuroscience and probability theory. He compared how often expressions often pointed out in AI-generated sentences are used, including “delve”, “intricate”, “underscore”, “showcase” and “nuanced”, before and after ChatGPT became widely popular in late 2022.

In computational linguistics, the use of those expressions rose about 7-fold to 14.0 times per 10,000 words in 2024 from 1.9 in 2022. Neuroscience also increased about 5-fold over the same period. By contrast, probability theory, used as a comparison group, showed only a small change to 1.3 from 0.9, and its confidence intervals overlapped. The researcher interpreted the increase in vocabulary as aligning with the degree of generative AI use by field.

Delve, long cited as a representative AI-style expression, appeared in 2.75 percent of the surveyed abstracts in 2024 but fell to 0.12 percent in 2026. The 2026 data, however, are an incomplete sample reflecting only up to roughly the first half of the year, so the final figure may change. The study said AI-specific expressions surged at one point and then partly declined as writers consciously avoided them.

The key point is that the appearance of a specific word does not mean the text was written by AI. The researcher also acknowledged that many abstracts since 2022 may have been drafted or polished using AI and then edited by a person. The analysis said that if AI expressions accumulate in writing under human names, detectors based on older human writing could increase false positives, while detectors trained on recent writing could accept an AI-influenced style itself as normal human expression.

The study is a preprint that analyses a limited genre of academic abstracts. It also does not test whether AI actually rewrites other AI text to bypass detectors. Rather than saying AI writing has completely changed human writing, it is closer to a warning that it is becoming increasingly difficult to distinguish human and AI writing using only word lists.