Imagine turning in an essay or report that you wrote, every word your own, and being told it is, in essence, fake—done with AI—not because anyone read your drafts or checked your notes with you, but because the piece had been sent through a widely used AI detection system, such as Turnitin, you may fail the class or get a hit on your next performance report.
An AI detector knows nothing about who wrote a piece. It measures predictability: how predictable each next word is, against the vast body of writing—the large language model—it was trained on. A composition that strives for common, safe, well-worn phrasing may score as machine-like. A score is a statistical guess about style, not a finding about authorship. It can prompt questions. On its own, it can never answer them.
So, a guess serves as a verdict. That is the whole problem, in one sentence.
Who gets hit hardest?
In a 2023 Stanford study, seven widely used detectors were run over 91 TOEFL essays by non-native English writers and 88 essays by American eighth-graders. On the eighth-graders, the detectors were close to perfect. On the TOEFL essays, the average false-positive rate was 61.22%. All seven detectors unanimously flagged 18 of the 91 essays as machine-written, and about 97.8% were flagged by at least one detector.
The researchers' explanation: writing in a second language tends toward common, safe word choices, which read as low surprise, as "machine." When the same essays were rewritten with richer vocabulary—same authors, same arguments—the false-positive rate fell from 61% to about 12%. Authorship did not change, only style.
That is not random error. That is patterned error—and it falls on the writers with the least room to sound "original" in a second language.
In a 2024 Common Sense Media survey of teens, about 10% said their work had been inaccurately identified as AI-generated—20% of Black teens, 7% of white teens and 10% of Latino teens. The survey reports what teens say happened; it does not claim detectors caused every flag. The gap could reflect tool bias, educator bias, or something else.
In a Center for Democracy & Technology survey of 460 public-school teachers, 68% said they used an AI detection tool regularly. That scale, plus acknowledged error rates and potential biases, is how a “problem” begins to look like an epidemic—especially to students!
AI cheating is not imaginary. UK university data, reported by The Guardian in 2025, showed almost 7,000 proven cases of AI cheating in 2023–24, or 5.1 per 1,000 students, up from 1.6 per 1,000 the year before. Those are proven cases, not just accusations—and they illustrate the pressure under which false accusations can happen.
Turnitin, for its part, reports reviewing over 200 million papers in its first year of AI detection, with about 11% flagged for at least 20% AI writing and about 3% for at least 80%. Those are detector flags, not findings of misconduct. At that volume, even a "small" error rate amounts to a large number of student-lives and careers potentially in jeopardy.
Here is the cruel double bind at the heart of it: Vendors sell certainty in the headline and disown it in the guidance.
Turnitin's instructor guide says its model " . . . may not always be accurate" and "should not be used as the sole basis for adverse actions against a student." Scores of 1% to 19% appear only as an asterisk, because false positives are more common in that band. Only qualifying long-form prose of at least 300 words is scored at all. An instructor or peer who treats a score as anything resembling a verdict is applying the tool in a way no vendor recommends.
Institutions have noticed. Vanderbilt disabled Turnitin's AI detector in August 2023, calculating that a 1% false-positive rate on its roughly 75,000 papers a year would amount to about 750 wrongly labeled papers.
By August 2026, Inside Higher Ed reported Yale, Johns Hopkins, and Indiana with policies banning or discouraging reliance on detectors as sole evidence, and at least a dozen universities, including Northwestern, Georgetown, and NYU, having simply disabled Turnitin's AI detection.
Newer tools have not closed the gap: a June 2026 comparison of four detectors on 160 known-origin texts found they missed most fully AI-written texts, and a large 2026 preprint testing 13 detectors on over 135,000 manuscripts by non-native writers found false-positive rates from 0.0% to 100%, depending on the detector. Which tool, which text, and which writer could change the answer completely.
Check that sequence again: the seller will not stand behind the score; nor will the schools . . . Only the students—alone, with failing grades.
The rules write themselves.
1. A probability is not proof. A detector estimates how much text looks machine-like. It cannot testify.
2. The burden stays with the accuser. Students should not have to prove they are human. The institution must prove misconduct, with evidence well beyond an AI detector score.
3. It’s the process, not the score. Drafts, notes, version history, prior work, sources, and a conversation with the student are evidence—slower to review, granted and more “expensive,” but the only kind that counts. A percentage can be a prompt for a conversation, nothing more.
4. A tool that flags some students more than others exaggerates the gap. Non-native English writers, formulaic academic writers, and students writing to a strict template are flagged more often. A fair process must account for that before it acts, not after.
5. A "folk tell" is not evidence. A dash, a tidy structure, a polished sentence—none of these is a finding. Punctuation is a keystroke, not a thought.
6. NEVER should peers be recruited to do what a university will not. If Vanderbilt, Yale, Johns Hopkins, and Indiana will not treat a detector score as sole evidence, a classmate's say-so should count for even less. (Some instructors now put peers in the detector's seat, a “folk tell” for which I have some anecdotal evidence, in fact—making this principle more urgent, not less.)
The classroom is where most first meet being judged by a score they cannot argue with—but it is not where this Faustian bargain ends. As Enrique Dans writes in Fast Company, we once outsourced memory to search, then navigation to GPS; now we hand off to AI the priceless labor of forming a judgment before expressing it.
What kinds of minds are these systems helping us form? And are we losing the ability to know how we came to know what we know—what is termed "epistemic dependence"—well beyond AI detection?
A score can be scrapped. A habit of letting machines judge for us (if a machine is working for us) is harder to overturn.
Your assignment: Take note. Keep the process. Keep the conversation. You be the judge . . .
Sources consulted:
How this post was developed: The topic and the angle are mine — this one grew out of a real encounter with an AI-detection accusation — and the final version, mine as well. My AI assistant (Muse) did the background research on the studies cited, drafted an early version, and recommended improvements in style and language. The studies and numbers are drawn from their published sources and public reporting, as listed in the references. Any errors in how those or the overall perspective on this issue are presented are mine.
No comments:
Post a Comment