Write the Non-Native Writer Clause as a Corroboration Rule, Not a Carve-Out

Write one sentence: no integrity finding rests on a detector reading alone, whoever wrote the paper. Three clauses are on the table. A carve-out exempts a named group from detector-based referral. A corroboration rule bars any finding that rests on a score alone. A course-level switch-off drops the reading where a cohort makes it unreliable. Only the last two avoid a roster of students, and only the corroboration rule keeps the detector and reaches the same students without labeling any.

Much of the demand traces to one paper. In the 2023 Stanford study, seven detectors read 91 human-written TOEFL essays and averaged a false positive rate of 61.22 percent ¹. At least one detector flagged 89 of the 91 ¹.

A December 2023 brief from the Center for Democracy & Technology connected that result to US civil rights law ². Title VI of the Civil Rights Act of 1964, a US federal statute, bars discrimination on the basis of race, color and national origin in programs receiving federal financial assistance ⁸. English learners are protected from discrimination under Title VI on both race and national origin grounds ².

Those two documents are a large part of why the topic reaches a policy committee. Neither supplies the clause.

The student-facing version of the same problem, and what a flagged writer can put in a reply, sits at what happens when a detector flags a non-native speaker.

A carve-out fails because it has to name its beneficiaries, and the exposure it answers does not stop at a name. The paper names non-native writers as the exposed group ¹; the reading that its second experiment reaches further than that label is this article's. The researchers rewrote those 91 human essays with a ChatGPT prompt for richer vocabulary, and the average false positive rate fell from 61.22 percent to 11.77 percent ¹.

A second prompt simplified the word choices in essays by US eighth graders and ran the effect backwards, from 5.19 percent to 56.65 percent ¹. The paper's reading is that the detectors penalized limited word choice, offered as a possible explanation rather than a proven one ¹.

Read the second half of that result slowly, because it is the part policy drafts skip. The eighth-grade essays came from a US dataset, and the paper treats them as native-speaker writing ¹. Plainer word choice alone moved the detectors' average false positive rate on their essays more than tenfold ¹.

The list is the first cost. Enrollment data does not record the range of English a student writes in. An institution that tries is likely to fall back on proxies such as visa status or language-program enrollment, which measure something else.

The record is the second cost. English learners are protected under Title VI on race and national origin grounds ². An exemption list would identify students by that membership and store it inside a disciplinary process, where people deciding cases read it. That inference is this article's, not the brief's; the brief's argument is about the effect of detectors on English learners, not about exemption lists ².

What that means for a particular district is a question for its own counsel. The drafter's point is narrower: the carve-out manufactures the record, and a clause covering every submission does not.

Coverage is the third cost. Writers flagged for plainness include students who are not English learners at all, which is why the same clause has to reach them too.

The parallel case, and the reason a disability-linked carve-out meets a sharper version of the same problem, is a policy for neurodivergent students and AI detectors.

A carve-out also has a shelf life. Written against one study and one generation of tools, it reopens the argument at every vendor release. A corroboration rule takes no position on vendor accuracy, so it doesn't expire.

The clause opens with one sentence: a detector reading may open a review and may never, by itself, support a finding of academic misconduct. That covers every student without naming any, which is why it survives the next accuracy release. The University of Pittsburgh's teaching center states the underlying reason in its own guidance, saying detectors are not reliable enough to act as proof ⁶.

The clause is only as good as its second half, which names what else counts. Define that list before the first case, because a definition written during a case is a definition written around one student. Four kinds of evidence are commonly listed:

• Timestamped draft history from the document the paper was written in
• Sources the student can produce and discuss in specific terms
• Prior work already on file in the same course, read for voice
• An interview whose questions are recorded and asked the same way of every student in the course

None of these is a detector score. Each still involves human judgment, and a committee should check that an interview or a reading for voice does not penalize fluency.

A corroboration clause is not a ban. Faculty keep the tool and keep the review it triggers. What moves is the burden. That distinction matters politically, because a committee that votes down a detector ban will still pass a sentence about what a finding requires.

Where that sentence has to land in the procedure, and who is allowed to hear the result, is worked out in designing an AI detection appeal process.

Ask three things in writing: which population the figure was measured on, how many documents it covers, and whether the set is published. Turnitin reports a false positive rate of 0.014 for English language learner writers against 0.013 for native English writers, on documents meeting its 300-word requirement ³. Pangram Labs reports no false positives on the 91 non-native samples of the 2023 study ⁴. Both numbers are real and neither closes the question.

A clean run of 91 documents says the rate is not large. It doesn't say the rate is near zero, and the distance between those two readings is the whole argument. The 91 essays are public, so any vendor can test on them, and a result on 91 documents bounds the rate loosely whichever vendor reports it.

The length floor is the second gap. Turnitin states its comparison for documents meeting a 300-word requirement ³. The weekly discussion post sits outside that figure. So does the short reflection.

Pangram's own model card also reports results by language. Its multilingual miss rate is 1.24 percent across 18 languages, and 5.3169 percent on Urdu ⁴. Those rows measure missed machine text in other languages, not wrongful flags on English written by non-native writers, so they are context rather than a third gap. What they do show, in the vendor's own evidence, is that performance varies with the language a writer works in. The same model card acknowledges a non-zero error rate and says false accusations of AI use can lead to serious consequences ⁴.

The fuller version of that list, written as procurement questions with the reasoning behind each, is ten questions to ask an AI detection vendor.

When the cohort and the assignment weaken the reading at the same time. Vanderbilt University disabled Turnitin's AI detector in August 2023, citing research that detectors are more likely to label non-native English writing as AI-written and saying Turnitin gave no detailed information on how its tool decides ⁵.

The University of Pittsburgh's teaching center went further and endorses no AI detection tool ⁶. Both were institution-wide positions. A course-level switch-off is the narrower version of the same decision, and at many institutions it can be made without a campus-wide vote; who holds that authority is a local question.

Vanderbilt set the 1 percent false positive rate Turnitin claimed at the time against the 75,000 papers it submitted in 2022 and said around 750 could have been incorrectly labeled ⁵. That sum assumes the vendor's rate holds on Vanderbilt's papers and treats every paper as human-written; both are assumptions. The same arithmetic on 1,600 papers gives 16 documents, on the same two assumptions.

The TeachAI toolkit goes as far as a ban. Its sample school guidance tells teachers not to use technologies that purport to identify generative AI use, on the ground that their accuracy is questionable ⁷. The same toolkit asks that any use of an AI system by a teacher or a student be disclosed and explained ⁷.

That trade is worth naming for what it is. Declaration replaces detection, and it runs on a disclosure culture a school has to build.

A targeted switch-off is narrower and easier to pass. The usual candidates are first-year composition with high international enrollment, intensive English programs, and timed in-class writing that falls under the vendor's length floor anyway.

Which institutions have published their own pages doing this, and the reasons each gave, is collected in universities that turned off Turnitin AI detection.

A switch-off is reversible at any point in the term. A disciplinary record built on a weak reading is far harder to undo.