AI Detectors and Neurodivergent Writers: Every Published Number (2026)

AI Detectors and Neurodivergent Writers: Every Published Number (2026)

Detection Drama Research · Last updated: July 29, 2026 · 9 min read

AI detector false positives for neurodivergent students are discussed constantly and measured almost never. Search the topic and you will meet the same statistic within three clicks: neurodivergent writers are flagged 3.2 times more often than their peers. It appears on vendor blogs, in AI-generated summaries, in legal marketing pages, and — until today — in an article on this site.

We went looking for the study behind it. There isn’t one.

This page is the provenance audit: every published number on AI detection and neurodivergent writing, each one traced to its source or explicitly marked as untraceable. Where a real measurement exists, we give it. Where the data simply does not exist, we say so instead of filling the hole with a number that sounds right.

1The One Study That Actually Measured This

Exactly one peer-reviewed study has tested the claim directly. Summer Chambers and Matthew C. Kelley of George Mason University published The Misclassification of Autistic Writing as AI-Generated at AIED 2025, and its design is the reason it matters: rather than surveying opinions, it ran a corpus through a detector and compared distributions.

Two details get lost when this study is summarised elsewhere. First, the authors explicitly noted that connections between the textual features of likely-autistic writing and the reported features of AI-generated text “were not straightforward” — the mechanism is not as clean as the popular explanation suggests. Second, they published no multiplier at all. Anyone quoting a specific “x times more likely” figure and citing this paper is citing something the paper does not contain.

The low absolute rate is easy to misread as reassurance. It isn’t. A sub-2% flag rate applied across every assignment a student submits over a four-year degree, by a tool used regularly in most classrooms, produces a meaningful chance of at least one accusation — and, as the research on false positive rates across detection tools shows, one accusation is all it takes.

2Where the “3.2x” Number Came From

Nowhere we can find. In July 2026 we ran five separate source searches for a primary reference behind the “neurodivergent students are flagged 3.2x more often” claim. Every result was a vendor blog, a secondary aggregator, or an AI-written summary — each citing the figure as established and none citing a study.

Why this matters practically: a student contesting a flag cannot use “3.2x” as evidence. The moment an academic integrity panel asks for the source, an unsourced multiplier becomes a liability rather than a defence — which is precisely when students discover the number they found online leads nowhere. The defensible material is the peer-reviewed corpus study, the institutional guidance describing the mechanism, and their own documented drafting process.

3What Is Measurable: Exposure, Not Accuracy

The false-positive rate for neurodivergent writers is unmeasured. Their exposure to detectors is measured, nationally, and the numbers are unambiguous. The Center for Democracy & Technology surveyed a nationally representative sample of 460 US public school teachers in grades 6–12.

Teachers reporting regular use of an AI content detection tool, 2023–24 school year. Source: CDT, Up in the Air (2024), Figure 4.

The teachers most likely to run a detector are the teachers whose students have IEPs and 504 plans. CDT’s prior-year survey also found that students with an IEP or 504 plan reported higher generative-AI use than their peers. Higher screening rate plus higher tool use produces compounding exposure — regardless of what any individual detector’s accuracy turns out to be.

4The Punishment Gap Nobody Scores

The most consequential number in the CDT data has nothing to do with writing at all. Forty percent of teachers agree that a student got in trouble for how they reacted when confronted about alleged AI misuse.

This is the mechanism that converts a statistical quirk into a disciplinary record, and it is the one place where being neurodivergent carries a documented, compounding penalty at the exact moment the detector score is being discussed. It also explains why detector-related anxiety shows up so heavily in student accounts: the confrontation is a second test, conducted without notice, scored on social fluency.

Teachers reporting students disciplined over generative AI, by detector use. Source: CDT, Up in the Air (2024).

Overall, 64% of teachers reported students facing consequences over generative AI in 2023–24, up 16 points in a single year. Among regular detector users the figure is 72%, against 48% for teachers who don’t use one. The tool does not merely observe discipline; its presence tracks with substantially more of it — a pattern consistent with what happens when a probability score is read as a verdict.

5The ESL Contrast: What a Real Rate Looks Like

It is worth seeing what a properly measured detector bias looks like, because the contrast explains what is missing here. For non-native English writers, the number exists and it is large.

Liang and colleagues ran 91 TOEFL essays through seven detectors and found more than half misclassified as AI-generated, while the same tools were near-perfect on essays by native-speaking US eighth-graders. That study is why detector bias against ESL students is treated as established fact rather than anecdote, and it is the reason universities can be pushed to act on it.

The proposed mechanism — low perplexity, meaning less lexical surprise from sentence to sentence — is the same one invoked for autistic writing. The difference is that somebody ran the experiment for one group and nobody has run the equivalent for the other. Neurodivergent students are arguing from a corpus study of Reddit posts and a set of teaching-centre explainers, while ESL students can point to a controlled comparison in a Cell Press journal. That asymmetry is the actual finding of this page.

6Risk-Exposure Checker

This estimates your exposure to detector screening from the published CDT figures. It does not estimate your chance of being falsely flagged, because — as established above — no such rate has been published for neurodivergent writers. Anyone offering you that number is making it up.

7The Legal Turn: Doe v. University of Michigan

In February 2026 a University of Michigan student filed the first US case we can identify that pleads an AI-detection accusation as disability discrimination under the Americans with Disabilities Act and the Rehabilitation Act.

The complaint’s specific allegation is the reason this case belongs in a data article rather than only in a legal roundup: it argues that “formal tone, meticulous structure, stylistic consistency, and heightened distress during oral confrontation” were interpreted as signs of dishonesty. That is a list of disability traits being read as evidence — the CDT reaction finding, restated as a legal claim.

For students, the practical consequence of the Michigan filing is procedural rather than statistical. A disability-discrimination framing shifts the question from “how accurate is this detector” — a fight nobody wins in a hearing room — to “did the institution follow a disability-informed process.” Earlier case outcomes suggest courts respond to procedural failure far more readily than to arguments about model accuracy, which is also the reasoning behind the growing list of universities that have disabled AI detectors outright rather than defend them case by case.

8Methodology

Frequently Asked Questions

Sources

Last updated: July 29, 2026 · Reviewed quarterly. If you can produce a primary source for the 3.2x figure, we will cite it and correct this page.