AI rewrites itself to evade detection, study of 19,000 abstracts finds

In computational linguistics, AI-associated word use rose from 1.9 to 14.0 instances per 10,000 words after late 2022, roughly a sevenfold increase. Neuroscience abstracts showed a similar pattern, climbing from 1.7 to 8.7 instances per 10,000 words, close to a fivefold rise. Mathematics abstracts barely changed, moving from 0.9 to 1.3 instances, which researchers treated as a control group; the study noted that overlapping confidence ranges made that small increase statistically indistinguishable from no change.

The tracked words included terms often linked to AI-generated phrasing, such as "delve," "intricate," "showcase," "nuanced," and "underscore." One tracked term, "delve," appeared in 2.75% of surveyed abstracts during 2024 before falling to just 0.12% within an incomplete 2026 data set. Because the 2026 sample remains partial, the reported decline may still shift once more data arrives.

Fırat Mıhcı, a computational linguist and founder of HumanizeMy.ai, released the findings as an open study meant to expose a blind spot in current AI detection methods used across many academic publishers. According to the study, a second AI model can rewrite an original AI draft until identifying traits become far less visible to reviewers. Once a human editor approves that rewritten draft, it gets published under a real name instead of being labeled as machine-generated.

The pattern creates a feedback loop in which AI wording gradually becomes accepted as ordinary human vocabulary. Detection tools trained on older writing samples risk falsely flagging real authors whose natural vocabulary now resembles AI-generated text more closely. Tools trained on newer papers instead risk absorbing AI-influenced writing as a normal baseline, making future AI text harder to catch. Editors and publishers reviewing these abstracts reportedly did not indicate that the wording carried any artificial origin before the manuscripts were accepted for publication.

"AI only needs its rewritten output to be accepted once as human. After that, the disguise becomes part of the answer key," Mıhcı said. The research does not claim every abstract containing these words was written by AI rather than a human author; it argues that AI-associated language has entered published writing in ways that make word-based detection increasingly unreliable over time. If AI-shaped phrasing keeps entering the published record undetected, the very baseline used to define human writing may keep moving.