It is becoming increasingly difficult to detect the use of artificial intelligence in scientific publications and academic texts. A recent study conducted by computational linguist Fırat Mıhcı showed that artificial intelligence models can completely evade detection systems by rewriting drafts they have generated themselves. The approval of these texts by human editors and their publication under real names is causing machine-generated language patterns to be adopted over time as part of the standard human vocabulary.
18 thousand 989 Academic Abstracts Examined in the Study
The study examined in detail 18 thousand 989 academic abstracts published in three different disciplines between 2019 and 2026. When the periods before and after the use of artificial intelligence tools that became widespread in late 2022 were compared, the rate of words associated with artificial intelligence in abstracts in the field of computational linguistics rose from 1.9 per 10 thousand words to 14.0. This increase was recorded as approximately sevenfold.
A similar trend was observed in neuroscience abstracts. The rate in this field rose from 1.7 to 8.7 per 10 thousand words, an increase of approximately fivefold. In mathematics abstracts, the rate rose from 0.9 to 1.3. The researchers treated this statistically uncertain change as a control group.
Studies such as the MIT report, which revealed that undergraduate assignments had previously been automated at educational institutions, had drawn attention to the impact of artificial intelligence on academic production. The new study demonstrates with concrete data the extent to which this impact is reflected in the published literature.
Changes in Frequently Used Words
The frequency analysis found that the words commonly used in artificial intelligence outputs changed over time. In the abstracts examined, the frequency of the word “delve,” used at a rate of 2.75% in 2024, fell to 0.12% in the incomplete 2026 dataset. On the other hand, it was determined that terms frequently used by artificial intelligence, such as “intricate,” “showcase,” “nuanced” and “underscore,” had become established in the literature.
It was determined that reviewers and editors had not made any note indicating that these texts were of artificial origin before they were published.
A Vicious Cycle in Detection Systems
The fact that artificial intelligence-generated texts correct themselves and come to resemble human writing creates a two-way risk for existing detection software. Detection systems trained on older texts may mistakenly flag original texts by real human authors influenced by artificial intelligence as AI-generated. By contrast, new models trained using current articles are inadequate at identifying artificial intelligence texts that will be created in the future, since they regard language shaped by artificial intelligence as a normal baseline.
The researchers point out that the reliability of text-based monitoring mechanisms has declined. Noting that the concept of human writing has changed over time, the researchers emphasize that methods going beyond stylistic analysis are needed to preserve academic integrity.