Do AI Detectors Flag Neurodivergent Writers More Often? What the Studies Say

Yes. In a 2025 peer-reviewed study, an AI detector flagged likely-autistic writers more often: 1.7 percent of their posts against 1.2 percent for a general group, at 300 words per post. It's the main AI detection false positive neurodivergent study, and universities warn about ADHD and dyslexia too.

These are AI text detectors, which score writing as human or machine-made. Tools that screen people for autism often come up in the same searches and are a different subject. These studies measure how often detectors flag text, with no claims about any diagnosis.

AI detection false positive neurodivergent research at a glance

A 2025 study of autistic writers provides the best data yet. It measured flag rates for one AI detector. University teaching centers also warn instructors about students with ADHD and dyslexia. A 2024 news test measured how spell check and rewriting tools affect AI scores. Other large studies of non-native English writers show how different detectors' results vary.

Group	Finding or guidance	Source
Autistic writers	1.7 percent flagged against 1.2 percent of a general group at 300 words, on one AI detector	Chambers and Kelley, AIED 2025
Writers with ADHD	Listed among students "prone to receive false positive ratings"	University of Nebraska-Lincoln teaching center
Writers with dyslexia	Listed in the same guidance	University of Nebraska-Lincoln teaching center
Neurodivergent students in general	"More likely to be falsely flagged"	Northern Illinois University teaching center
Writers who use editing tools	Spell check and grammar suggestions barely moved AI scores, and whole-text rewrites moved them a lot	Bloomberg Businessweek test, 2024
Non-native English writers	False positive rates from 0.0 to 100.0 percent across 13 detectors	Park, Jeong and Kim, EMNLP 2026

About two thirds of teachers say they use an AI checker regularly, according to a survey of more than 450 instructors. The Center for Democracy & Technology published it in March 2024. Even a low error rate can add up over the number of papers students write each year, Bloomberg Businessweek reported.

Autism AI detector study: what the authors found

The paper is called "The Misclassification of Autistic Writing as AI-Generated" by Summer Chambers and Matthew C. Kelley from George Mason University. It's peer reviewed, and Springer published it in Artificial Intelligence in Education (AIED 2025). You can get a free copy on arXiv. The researchers write that the anecdotes about autistic writers being misidentified more often "have not been formally investigated." They wanted to test these claims.

Their data set consists of about 60,000 Reddit posts, one post per user. They excluded short posts (under 1,000 characters). The likely-autistic group came from 13 subreddits about autism. The general group came from 12 popular subreddits chosen to be similar in style, with long first-person narratives on mental health, social situations and advice. General posts with autism keywords were removed.

The posts were obtained from the Reddit API and covered posts from about 2021 to 2024, as well as a collection of older posts (from 2010 to 2020). Posts including both autism keywords and phrases such as "my daughter" or "my nephew" were also removed to keep posts written by autistic writers.

The authors call the first group "likely-autistic" because no diagnosis was checked, and anyone can post in any subreddit. Some autistic people post in the general subreddits too, so the authors suspect the real gap between autistic and non-autistic writers may be larger than the one they measured.

The detector was OpenAI's RoBERTa GPT-2 detector, released in 2019 and free to download. A post got a "Fake" label, which counts as a flag here, when the detector put its AI probability above 0.5. In the first experiment, shorter posts were flagged more often than longer ones, and the likely-autistic posts were shorter on average. Length could have explained part of the gap, so for the second experiment the authors kept only posts over 300 words and cut each to exactly 300.

Measure	Experiment 1	Experiment 2
Posts analyzed	59,947	33,216
Text used	Posts of 1,000 characters or more	Posts cut to exactly 300 words
Likely-autistic group flagged	1.9 percent	1.7 percent
General group flagged	1.5 percent	1.2 percent
Adjusted odds ratio	1.246 (p = 0.001)	1.501 (p below 0.001)
Higher perplexity	Lower chance of a flag	Not significant

The gap stayed when length was equal. In Experiment 2, 1.7 and 1.2 percent are about 17 against 12 posts per 1,000, so most posts in both groups were not flagged.

They also report odds, taken from a regression (statistical model) controlling for sentence length, word length, perplexity, burstiness and post date. They talk about "a 25% greater chance" of a flag in Experiment 1 and "a 50% greater chance" in Experiment 2. These are relative numbers on a small base. The difference in Experiment 2 is half a percentage point in the raw percentages; use the percentages if you want to know how many flagged posts there were.

Why detectors may flag neurodivergent writing

University guidance and researchers point to word choice. "The Challenge of AI Checkers", a guide for instructors from the University of Nebraska-Lincoln, says there's no one reason neurodivergent students receive false positives. It says they're often related to reliance on repeated phrases, terms and words. The same guide says the lack of pronouns such as I, me and we can be misconstrued as AI writing, depending on the checker.

A Bloomberg Businessweek report from October 2024 says the students most susceptible are "likely" those who write in a more generic manner, including neurodivergent students. The reporters heard this from students, academics and AI developers. James Zou of Stanford was senior author of the 2023 study on non-native writers. The same report quotes him: "If the word choices tend to be more generic and formulaic, that work has a higher chance of being flagged by AI detectors."

In the autism study, a higher rate of flags was associated with no reduction in perplexity or burstiness. Both scores were nearly identical for the two groups, but tended to be slightly higher for the likely-autistic group. The authors note that relationships between features of text from authors who are likely to have autism and those from AI-generated text "were not straightforward".

Our guide to how AI detectors work explains perplexity and burstiness if they're new to you.

ADHD writing flagged AI: what universities tell instructors

University teaching centers have begun alerting instructors about students with ADHD. The Nebraska guide says: "Students that are neurodivergent (autism, ADHD, dyslexia, etc...) are also prone to receive false positive ratings". Northern Illinois University's "AI detectors: An ethical minefield", from December 2024, says: "Neurodiverse students are also more likely to be falsely flagged for AI-generated writing".

Whitney Gegg-Harrison and Claire Quarterman wrote a 2024 book chapter in Academic Integrity in the Age of Artificial Intelligence (IGI Global). Its small corpus study found much higher than advertised false positive rates across popular detection tools. Drawing on that study, news reports and student testimony, the chapter "presents the possibility that neurodivergent writers, along with L2 writers, are more likely to be impacted by false positives". L2 means second-language.

One finding from the autism study applies to anyone who writes short pieces, such as discussion posts. In Experiment 1, shorter posts were flagged more often than longer ones. We cover what detectors say about length in our post on the AI detector minimum word count. Saving your drafts and notes as you write is a good way to have a record of your writing process. This answers a flag better than any score.

Dyslexia AI detection and editing tools

The Nebraska guide includes students with dyslexia in its warning about false positives. Many writers use spell check, grammar suggestions or rewriting tools as they draft, and two studies show how those tools change AI scores.

In a 2024 Bloomberg Businessweek test, a popular writing assistant's spell checker and grammar suggestions had only a marginal impact on making documents appear more AI-written. Whole-text rewrites were different: asking the same assistant to "improve" an essay or "make it sound academic" turned work scored as 100 percent human-written into 100 percent AI-written.

Hyeonchu Park, Gahye Jeong and Bugeun Kim studied professional editing in "Style as a Confound", presented at EMNLP 2026. They used 135,389 pairs of documents from an academic English editing service, from 2018 to 2025. Each pair is a non-native writer's manuscript before and after a native-speaker editor edited it. Across 13 detectors, the same edits raised scores on some detectors and lowered them on others, and score changes grew with the amount of editing.

They then broke the edits down by type on one of the detectors, an OpenAI RoBERTa model. Grammar edits produced a small negative shift in its score, while syntactic edits (which change sentence structure) produced the largest positive shift. They conclude that detector outputs should not be treated as definitive evidence of AI use, especially where editing is common. If you use a tool that rewrites whole passages, keep your earlier versions and check what your course allows.

False positive rates depend on the detector

The false positive rate varies with each detector. The rate is for the particular detector that the study was testing. Liang and colleagues' 2023 study in Patterns found almost perfect accuracy on essays by US eighth graders. In contrast, the detectors marked more than half of TOEFL essays by non-native writers. See our post on detector bias against non-native writers for the details.

Our working paper on non-native writers and false positives, from TextPulse Research in 2026, scored 5,600 ICNALE essays written before modern AI tools. One research classifier, a logistic regression, flagged 2.7 percent of learner essays and 3.3 percent of native essays. A gradient boosting model trained on the same features and essays flagged learners at 12.5 percent and natives at 3.8 percent. So whether one group gets flagged more often depends on the detector.

In the Park study, the 13 detectors' false positive rates varied widely for human-written texts, ranging from 0.0 percent to 100.0 percent. Two prominent detectors were tested by Bloomberg Businessweek on a random sample of 500 college application essays sent to Texas A&M in summer 2022, before ChatGPT. The services incorrectly classified 1 percent to 2 percent of the essays as likely AI.

In the same report, Turnitin said its document-level false positive rate was below 1 percent and its sentence-level rate was 4 percent. So the figure that matters for a student is the one for the detector their school uses.

If you are flagged, or if you review a flag

For students

Your draft history is where you start. It shows earlier drafts and notes on how your text developed, and a score can't. Ask in writing for the full report and for the policy that applies. The steps are covered in our posts on what to do first if you are accused and on writing an AI detection appeal letter.

The best response to a flag is to show your work. Keep a record of what you wrote. Use Google Docs or Microsoft Word (with version history enabled) to draft and keep track of your notes, outlines and sources you read. If you use assistive technology like spell check, dictation or text-to-speech, you can tell your instructor or accessibility office which ones you use and how.

You might also want to contact your university's disability or accessibility office. Decide if you want to tell your teacher or panel about your diagnosis. You could say something like this: The rates vary from detector to detector. A peer-reviewed study showed a higher flag rate for likely-autistic writers on one of them. Your drafts are more important than research on groups because research on groups can't show how your own essay was written.

For staff

The score is just an estimate. It's a reason to ask about the writing process. Ask to see drafts and notes if they exist. Ask the student how they wrote the piece. A diagnosis or a writing style isn't evidence for or against AI use. Ask which tools the student used, like spell check or dictation. Spell check and grammar suggestions barely moved AI scores in Bloomberg's 2024 test, but whole-text rewrites moved them a lot. The same Bloomberg report says detector makers say their scores are one data point to guide a teacher. Chambers and Kelley recommend further critical examination of these models and their use in academic contexts.

Tools on this site

The free perplexity checker estimates the word variation in your draft, and the free burstiness checker charts your sentence lengths. Neither number proves anything about who wrote the text. The TextPulse AI humanizer is designed to lower AI scores on a draft you wrote yourself while preserving citations and terms. If you use it, check that your course allows such a tool, and read the output. Nothing in this research says you need to change how you write.