AI’s writing fingerprints are becoming part of human prose

AI-generated writing may be getting harder to detect for a reason more consequential than improved paraphrasing: its vocabulary is entering the published record and becoming part of what reviewers recognize as normal human prose.

An analysis of 18,989 academic abstracts spanning computational linguistics, neuroscience and mathematics found that several words associated with AI-assisted writing became substantially more common after widely used generative-AI writing tools emerged in late 2022. The change was strongest in fields where researchers routinely use language models for drafting or editing. Mathematics showed little movement and was used as a control group.

That creates a problem for detection systems built around style. If machine-influenced wording is repeatedly accepted under human authors' names, future detectors may learn that wording as evidence of ordinary authorship rather than evidence of generated text.

Recommended reading

HKUST’s 7G plan makes AI agents part of the network

Sophia Reynolds • Sep 1, 2026 • 5 min read

The vocabulary shift was uneven across fields

The study compared word-frequency patterns before and after late 2022. Its reported figures show a large increase in tracked AI-associated vocabulary in computational linguistics and neuroscience, but not in mathematics:

The mathematics result was used as a comparison group. Its confidence ranges overlapped, meaning the small increase was statistically indistinguishable from no change across the full period. The result doesn’t show that mathematicians avoided AI tools; it shows that this particular vocabulary signal changed far less in that corpus.

The researchers also tracked individual terms, including “delve,” “intricate,” “showcase,” “nuanced” and “underscore.” “Delve” appeared in 2.75% of surveyed abstracts during 2024, then fell to 0.12% in the incomplete 2026 data set. That decline could change as more 2026 abstracts are added, so it isn’t evidence that the term has permanently disappeared from AI-shaped writing.

The more durable finding is the broader shift in frequency. A single word can become a weak signal once authors, editors and models all encounter it more often. Word-based detectors face a moving target rather than a fixed distinction between human and machine prose.

Why rewriting can defeat a detector

The analysis describes a feedback loop involving at least two language-model passes. One model produces an initial draft. A second model then rewrites it to reduce the stylistic features that detection systems are looking for. If a human editor approves the result, the text can appear in a published paper under a real author’s name without being labeled as machine-generated.

That process doesn’t require the rewritten text to be perfectly human in any philosophical sense. It only needs to pass review once. After publication, it becomes part of the corpus used to define contemporary academic writing—and potentially part of the material used to train future detection systems.

“AI only needs its rewritten output to be accepted once as human. After that, the disguise becomes part of the answer key.”

The claim is narrower than “these words prove AI wrote an abstract.” The study does not establish that every abstract containing the tracked vocabulary was generated by a model. Human writers can independently choose the same words, and editors can introduce them during normal revision. The finding is that the aggregate frequency pattern has changed enough to make vocabulary alone a weaker basis for deciding who—or what—produced a passage.

That distinction matters for publishers. A detector trained on older writing may increasingly flag legitimate authors whose vocabulary now resembles AI-assisted prose. A detector trained on newer papers may absorb machine-influenced writing as its baseline and become less sensitive to later generated text. Either way, the accepted literature can shift the reference point used by the detector.

The study leaves a measurement problem

The data cover abstracts published from 2019 through 2026, but the 2026 sample is incomplete. That makes the exact trajectory of individual words provisional, particularly the reported fall in “delve.” The supplied findings also don’t establish a universal AI detector, a false-positive rate, or a causal link between each vocabulary increase and the use of a language model.

What they establish is a warning about relying on surface style as a stable forensic signature. The strongest evidence here is the contrast between fields: computational linguistics rose from 1.9 to 14.0 instances per 10,000 words, neuroscience followed a similar pattern, and mathematics barely moved. That pattern is more informative than any list of supposedly “AI words.”

For editors, authors and institutions, a clean result from a writing-style detector may mean only that the text matches the current published norm. As AI-assisted language continues to enter that norm, the disguise doesn’t just evade the test. It helps define what the test considers human.