AI detection in academic writing can create a problem that receives far less attention: researchers may rewrite legitimate work simply to satisfy an algorithm. In a 2026 letter published in Annals of Biomedical Engineering, Louie Giray calls this practice “un-AI-ing.” The term describes deliberate changes made to a text to lower the likelihood that an AI detector will flag it.
Giray’s central point is that the same detection system can encourage two very different behaviours. One conceals inappropriate AI use. The other pressures honest researchers to make clear writing less clear. That distinction gives editors, universities, and researchers a more useful way to think about the effects of AI detection.
When researchers begin writing for the detector
Scientific writing depends on consistency. A methods section often follows established reporting conventions. Technical terms need to remain stable, and a well-edited argument usually avoids unnecessary variation. Yet AI detectors commonly assess statistical features such as perplexity, which concerns how predictable the wording is, and burstiness, which concerns variation in sentence structure and length.

The conflict is easy to see. Some qualities valued in academic writing resemble the qualities a detector associates with machine-generated prose. A researcher whose manuscript receives a high AI score may begin replacing precise terms, breaking up coherent sentences, or introducing stylistic irregularities. The detector then starts shaping the prose, even when AI did not generate the original argument.
I have discussed related reliability problems in my earlier review of AI detectors in education. Giray adds a behavioural dimension to the issue. A false positive does more than misclassify a document. It can change how people write and encourage them to optimize their work for a score that does not establish authorship or misconduct.
Two forms of un-AI-ing
Giray’s article is a conceptual letter and does not report an original empirical study. He draws on previous research about detector accuracy, bias, academic publishing, and AI-assisted writing to develop a taxonomy of un-AI-ing. He then considers what it could mean for scientific publishing, particularly biomedical research.
The first form, which he calls opportunistic un-AI-ing, involves disguising AI-generated content. A writer may use a humanizer, repeated paraphrasing, or other alterations to make synthetic text harder to detect. These changes conceal how the text was produced without addressing whether its claims, sources, or analysis are trustworthy.
The second form is compliant un-AI-ing. Here, a researcher changes human-authored or legitimately assisted writing after a detector flags it. Giray argues that this burden may fall especially heavily on multilingual researchers whose formal and consistent academic English can be misread as evidence of AI generation.
The article brings earlier studies and examples together in support of this argument. It does not directly measure how frequently either form of un-AI-ing occurs. The concept is persuasive, but we still need systematic research into its prevalence and effects.
What journals and universities could do instead
Giray recommends moving away from automated gatekeeping and toward policies based on disclosure, context, and human judgement. This aligns with the argument I explored in my review of AI disclosure in academic publishing: disclosure becomes more useful when policies distinguish between substantive generation and ordinary assistance such as proofreading or language editing.
For editors, a detector score should never carry the weight of proof. A more defensible review process would check whether the references exist and whether the evidence supports the conclusions. Reviewers could also examine the disciplinary detail in the methods and ask authors to describe consequential uses of AI. These checks address the integrity of the research itself.
Universities face a similar decision. If students or researchers learn that polished prose attracts suspicion, they may start managing detector scores instead of improving their arguments. This can turn academic integrity into a performance of statistical “humanness.”
Clear policies should state what forms of AI assistance are allowed, what must be disclosed, and how concerns will be investigated. Authors also need a meaningful opportunity to explain their process and provide drafts, notes, sources, or other evidence of authorship.
The article also fits Sarah Eaton’s idea of postplagiarism, where hybrid human-machine writing has become part of the scholarly environment and older binary categories no longer explain every case. Giray’s two-part taxonomy helps locate the ethical issue more precisely.
Opportunistic un-AI-ing hides a relevant use of AI. Compliant un-AI-ing may be a defensive response to an unreliable accusation. Treating both as the same behaviour would erase a significant difference in intent and responsibility.
For me, the most useful part of the paper is its shift in the evaluative question. A detector asks whether a document’s statistical pattern resembles AI-generated text. Research integrity asks whether the author can account for the ideas, evidence, methods, and decisions in the document. Those questions are not equivalent.
Readers interested in related issues can find more reviews in my AI in Academic Research collection.
Full reference
Giray, L. (2026). Un-AI-ing: Compliance, evasion, and the distortion of research writing in the age of AI detection. Annals of Biomedical Engineering. Advance online publication. https://doi.org/10.1007/s10439-026-04242-2




