No Published Study Located as of September 2026 Measures Detector Bias in Neurodivergent Writers: Write a Consultation Step Into the Policy

One named case, and a magazine's characterization of a wider group. Bloomberg Businessweek reported in 2024 that Moira Olmsted, a student at Central Methodist University, was flagged in 2023 and told the university she has autism spectrum disorder and writes in a formulaic way a detector can misread ¹. Her grade was changed, and she was still warned that another flag would be treated the way plagiarism is treated ¹.

A search of the published literature in September 2026 for a measured false-flag rate on autistic, ADHD or dyslexic writers found none: no controlled study, no published vendor audit, no replication of the non-native-speaker work on this cohort. That thinness is the finding. An absence found by searching is a dated search result, and it is offered here with its date attached.

The same reporting names the wider group. It does not measure it. The students most susceptible to an inaccurate accusation are likely those who write in a more generic manner, either because they are neurodivergent, speak English as a second language, or learned to use more straightforward vocabulary ¹. Who said so matters. It came from the magazine's sources rather than from any study, and it is not a measurement.

Bloomberg also reported that Turnitin does not train specifically on neurodivergent student data and has no access to medical histories to assess that classification ¹. The same passage carries the company's own position. Turnitin's vice president for AI told Bloomberg that the company intentionally oversamples underrepresented groups in its data set. Internal tests, on the same account, have shown its model does not falsely accuse students writing in English as a second language, and its overall false positive rate for entire documents is below 1 percent and improving with each release ¹. A district weighing the tool is entitled to both halves.

Compare what exists one category over. For writers working in a second language there is a measured figure. A 2023 Stanford study ran 91 human-written TOEFL essays through seven detectors and recorded an average false positive rate of 61.22 percent ². Having ChatGPT rewrite those same essays with richer vocabulary cut the rate to 11.77 percent ². Near-perfect accuracy on essays by US eighth graders came out of the same run ².

That number is not transferable. It came from TOEFL essays run through seven named tools in 2023, and lending it to a different cohort would be an assumption nobody has tested. What that study supports, and what it does not is a separate question.

Licensed special education teachers, by a 2024 survey from the Center for Democracy and Technology. In a nationally representative survey of 460 US middle and high school teachers fielded in late 2023, 68 percent said they use an AI content detection tool regularly, up 30 percentage points in a year ³. Licensed special education teachers reported using one more often than their colleagues, 76 percent against 62 percent ³.

What that does not measure is whose work gets scanned. A licensed special education teacher does not teach only students with plans, and the survey asked teachers about their own practice. The inference, offered as an inference, is that a tool with no published error rate located for neurodivergent writers is used regularly by more of the teachers licensed to serve students with disabilities.

The survey carries something else worth taking into a policy meeting. Teachers who use a detector regularly were more likely to report students getting in trouble over AI, 72 percent against 48 percent of teachers who do not use one ³. Which drives which, the survey does not show: a school already worried about AI use may be quicker to buy a detector. A district can still ask the design question underneath the correlation.

None of that measures bias. It measures how often teachers reach for the tool, and that is the half an administrator controls. A district can't fix a model it did not train. What it can decide is where the reading enters the process, who sees it first, and what the reading is permitted to do on its own.

Not the flag, and not the score. What changes is the consequence, and who has to be in the room before one lands. In the United States, for a K-12 student with an IEP, a removal of more than 10 consecutive school days, or a series of shorter removals that form a pattern, is a disciplinary change of placement ⁸. That triggers machinery already written into federal regulation ⁴.

The regulation sets a clock. Within 10 school days of any decision to change the placement of a child with a disability over a code of conduct violation, the district, the parent and relevant members of the child's IEP team have to review the conduct ⁴. They settle two questions: whether the conduct was caused by, or had a direct and substantial relationship to, the child's disability, and whether it was the direct result of a failure to implement the IEP ⁴.

Read that against an AI accusation and the awkwardness shows. A manifestation determination asks whether a disability explains the behavior. Here the disputed behavior is a writing style, and the evidence that it happened at all is a percentage from a model with no published audit on this cohort. The team is being asked to adjudicate causation for an act that has not been established.

IDEA's manifestation determination binds school districts under Part B ⁴. Non-discrimination duties run wider. Section 504 of the Rehabilitation Act bars exclusion or discrimination against a qualified individual with a disability in any program receiving federal financial assistance. Title II of the Americans with Disabilities Act covers state and local government entities ⁵. Title II requires reasonable modifications to policies, practices and procedures where needed to avoid discrimination, unless the change would fundamentally alter the service ⁵. The Department of Education's guidance for postsecondary students states that practically every school district and postsecondary school in the United States is subject to one or both laws ¹⁰.

One difference between the two settings matters more than either statute. A college does not identify a student as having a disability: the student does, and only when asking for an academic adjustment ¹⁰. So a district clause and a college clause cannot be the same sentence.

A district or college could treat consultation as a reasonable modification of this kind. Whether the law requires it is a question for counsel, and so is how any of this applies in a given district. Consulting the office that already holds the documentation, before a finding rather than after an appeal, alters no academic standard and removes nothing from the process.

What an appeal actually turns on is procedure and record, which is why the record has to start earlier than the appeal does.

Two clauses, one per setting. In a district, when a flagged student has an IEP or a 504 plan, the case manager or 504 coordinator is consulted before an integrity report is filed, under the confidentiality rules the district already runs. In a college, the first conversation tells the student that disability services can be brought into it. Nothing is confirmed to anyone without the student's consent.

That is a sequencing rule rather than a shield. It decides no case and exempts nobody. What it does is put the person holding the documentation into the conversation while the conversation can still change.

The split is not pedantry. IEPs and 504 plans are K-12 instruments held by the district. In the United States, for a student served under IDEA, 34 CFR 300.622 requires consent from the parent, or from a student to whom those rights have passed at the age of majority, before personally identifiable information goes to parties other than officials of participating agencies, unless the disclosure is one FERPA's regulations at 34 CFR part 99 allow without consent ⁹. That rule does not reach a student who has a 504 plan and no IEP, and which of its paths applies to an integrity officer is a question for district counsel. A college does not write or hold an IEP for its students, and disclosure there is the student's own choice ¹⁰.

One public university system's code shows where such a clause can sit. The City University of New York asks a faculty member who suspects a violation to review the facts and circumstances with the student first, wherever that is feasible, before filing a report ⁶. A consultation step sits beside that conversation.

One row there is easy to get wrong. A diagnosis does not belong in an academic integrity file, and neither does the bare fact of a plan. The useful part of a consultation is the advice, and the question it answers is whether the writing pattern under suspicion is one the office recognizes.

An oral follow-up is one more piece of evidence, and it is the step a plan is most likely to bear on. A student can have written a paper and still struggle to discuss it on the spot. Consult before it, and adapt its format where the plan calls for that. What an oral follow-up can and cannot settle is worth reading before one is scheduled.

Borrowed wording saves an argument. Human's standing caveat on a result reads: This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A district can lift those two sentences into its own procedure as the limit on what a reading may be used for.

A district drafting this for the first time can start from an institution-level clause and cut it down.

Ask whether the error rate has been measured on writers with disclosed disabilities. Bloomberg reported in October 2024 that Turnitin did not train specifically on neurodivergent student data and had no access to medical histories to assess that classification ¹. Whether that still holds, for Turnitin or for another vendor, is the question to put in writing. An answer of no is usable. An evasion is the finding, and it belongs in the procurement file next to the price.

Ask, too, whether the tool can be tested by somebody outside the company. Bloomberg reported that Turnitin declined to make its service available for testing ¹.

One multiplication follows from the vendor's own published figure. Vanderbilt University ran it in August 2023, against Turnitin's stated 1 percent false positive rate and its own submission volume. That was 75,000 papers in 2022, around 750 of which could have been incorrectly labeled as carrying some AI writing ⁷. Vanderbilt switched the detector off and said publicly that it does not consider AI detection software an effective tool ⁷.

Any district can do that arithmetic in a spreadsheet with numbers it already owns. Papers submitted last year, times the vendor's published rate, gives a rough count of papers a purely automatic process could have labeled wrongly. That is an upper-end sketch resting on two assumptions, that every paper was written by a person and that the vendor's rate holds on this district's writing, and neither has been measured.

Human has no measurement on this cohort either. Its published figures come from held-out academic writing, and they are in-house measurements on its own sets rather than an independent audit. None of them says anything about autistic or ADHD writers. Saying so is better than borrowing a figure measured on a different population, which is the specific error this page is written against.