Can a corrected text look more artificial to a detector?
A 2026 study examines what happens when a text’s style changes while its authorship remains the same. It invites a careful reading of detector percentages.

Cover image generated with AI for Ed-IA.
The development worth examining
Park, Jeong and Kim compare 135,389 pairs of academic English manuscripts before and after professional editing, using thirteen detectors. Editing increases AI scores for some tools and reduces them for others. Style can affect the output even when authorship and content remain unchanged. [1]
The question this raises
If a result changes after editing, we need to understand what the tool measures. An indicator of style cannot automatically establish who wrote a piece of work. This matters when the output appears as a precise percentage. Decimal places convey apparent certainty; interpreting the result requires an independent justification.
What we cannot conclude
The sample concerns professionally edited academic English. It does not establish a false-positive rate for Catalan or Spanish school writing. Nor does it establish that all detectors behave alike or that every correction increases suspicion. The consulted version is on arXiv; its record states presentation at EMNLP 2026. [1]
| Claim | Careful interpretation |
|---|---|
| Style can affect the output | Check what the score represents |
| Different tools respond differently | They do not share one error percentage |
| The sample is academic English | Specific evidence is needed before applying it to school work |
An implication for assessment
Ed-IA’s proposal is to investigate the process before attributing authorship. Discussing decisions, reviewing requested drafts and asking for a brief applied explanation can provide information about learning. None of these indicators resolves every case alone. They should be accessible, known to students and planned as part of the task.
For example, students could justify a change between versions and explain why they rejected an alternative. The discussion should allow them to clarify doubts and provide context. This is a teaching proposal to adapt to the school, rather than a finding measured by the researchers.
Reading the next announcement
When a new detector appears, ask which texts were used to test it, what errors it makes and whether it was validated for the students’ language and age. Distinguish a commercial claim from an independent evaluation. If that information is missing, the percentage tells us less than its presentation suggests.
This study helps formulate a practical question: what evidence would change our view of a student’s work? Deciding that before a difficult case arises helps prevent an improvised judgement based on a number.







