A judgement about a piece of work should rest on what the work says and what it is worth. When it rests on a machine's estimate of who made it, the judgement lands on the person, and the person has no way to answer it.
A piece of work can be asked two questions. One is about the work: what does it say, and does it hold? The other is about its maker: did a machine write this? In classrooms the second question has started to stand in for the first. Koher's tools ask only the first, and this position says why.
The left question can be settled by anybody who reads the work. The right one cannot be settled by reading at all.
A percentage that carries a penalty
Indian higher education already runs one judgement on a percentage. The University Grants Commission's regulations of July 2018 sort a thesis or dissertation by how much of it matches other text: up to 10% carries no penalty, above 10% the student revises, above 40% they are barred from resubmitting for a year, and above 60% their registration is cancelled.
That percentage can at least be opened. A similarity report shows the passage that matched and the source it matched, and a student can look at both and say what happened. The AI-writing score that the same software now gives cannot be opened in that way. It highlights the sentences it suspects, and there is no source behind a highlight to check it against. Turnitin's own guide says its model "may not always be accurate" and "should not be used as the sole basis for adverse actions against a student". Since July 2024 it has shown an asterisk in place of any score under 20%, because false positives are more frequent there.
A match can be checked against its source. An estimate of authorship has no source to check it against.
How good the detectors are
In 2023 a test of fourteen detectors, Turnitin among them, concluded that they were "neither accurate nor reliable" (Weber-Wulff and colleagues, International Journal for Educational Integrity). The same year, detectors were found to misclassify writing by non-native English speakers as machine-made (Liang and colleagues, Patterns).
Later tests are kinder on one point. A 2026 study of four detectors found that all of them correctly identified fully human texts, and still concluded that they "should not be used as sole evidence in high-stakes decision-making" (Van Vlasselaer, Van Droogenbroeck and Spruyt, in the same journal). The maker of one of them, Pangram, reports a false positive rate of 0.0041%, "roughly 1 in 24,000". A reviewer at PCMag, testing that tool in September 2026, got it to miss machine-written sentences he had added to his own draft, and found that of the two sentences it did flag, one was his.
So the accuracy is disputed, and it changes with every new model on both sides. Nothing below depends on how that dispute ends.
Even an accurate score
However good a detector gets, a small error rate is still a number of people. When Vanderbilt University switched Turnitin's detector off in August 2023, it did the sum itself: 75,000 papers a year at the vendor's stated 1% meant about 750 flagged in error. Washington State University, cancelling its contract in February 2026, counted 148,547 assessments in one term, or almost 1,485 at the same rate. It also reported that a third of its hearing-board cases about AI between 2023 and 2025 ended in a finding of not responsible because the detector's report was the only evidence. The score had been enough to bring each of those students before a board.
The rate is the vendor's figure and the papers are the university's count. Neither is Koher's.
A student who is flagged is asked to show that they did not use something, and the report gives them nothing to point at.
A detector that was right every time would still say nothing about the work itself: whether the argument holds, or whether the drawing solves the problem it was set.
A number posing as evidence
Not Counted, Not Named says that a figure replaces a possibility, and that everything afterwards argues from the figure. Called a Machine adds that a number can pose as evidence of something else. An AI score works in that way: a statistical estimate about a text gets read as a fact about a person.
The same judgement, without the tool
People make this judgement unaided, and in both directions across the classroom.
Teachers gave essays lower marks once they were told a machine had been involved, and made assumptions about what the student could write (Tan, Wang and Xu, Applied Linguistics, 2025). Writers suspected of using AI were rated lower and were less likely to be hired (Kadoma, Metaxa and Naaman, 2025). Students do it too: 472 of them rated feedback lower once they were told an AI had written it, and saw the AI as the less credible provider (Nazaretsky and colleagues, Journal of Computer Assisted Learning, 2025).
Nor do the accusations follow the text. In a study of 25 million forum comments, the features that actually distinguish machine prose did not predict which human writing got accused of being it (Miklian and Katsos, 2026, a preprint).
One thing the accusation can follow is the reader's estimate of the writer. Work that is better than the reader expected of that person is the work most easily put down to a machine. Good work used to be how an underestimated person corrected the estimate. Read this way, it stops counting for the people who most need it to count.
The yellow box is the only evidence the person has, and the red box is what removes it.
This is the loss Called a Machine describes, arriving through assessment. A person whose work is set aside as a machine's has had more than a mark taken away. They have been told that what they made cannot have been theirs.
What Koher does
Koher's tools read the work. They describe what a piece says and how it holds together, the judgement sits in code anyone can read, and a person decides. Koher builds no detector. Nothing it makes tries to guess whether a person or a machine wrote something.
Where authorship does matter, the finished text is the wrong place to look for it. Split-Domain Cognition's own account of the split ratio says it "can only be claimed from recorded provenance". How much of a piece a person made is something recorded while it is being made: dated versions, the drafts, a note of what was asked of a machine and what was kept. Afterwards there is the person, who can be asked about the work and can answer for it.
Two questions, and a separate way of answering each. The detector's score is on neither line.
Where this stops
It does not say that machine help is nothing, or that a course may not forbid it. Where it is forbidden, breaking the rule is a matter between a teacher and a student, and it is settled by the record and a conversation.
It does not say a detector is never right. Counting across thousands of documents is a different use from judging one person's work, and this position is about the second.
And it describes how Koher builds, and what it will not build. Other people set the rules for their own classrooms.
A mark is a statement about a piece of work. Once it depends on a guess about what kind of thing produced the work, it has become a statement about the person, made by a means they cannot question. That is the reason Koher's tools read the work and leave the question of its maker to the record and the person.