AI detectors and neurodivergent writers: what the evidence shows so far

The short answer

There is now real evidence — case reports and a first empirical study — that AI detectors flag neurodivergent writers' genuine work at elevated rates. The evidence base is younger and thinner than the well-measured bias against non-native English speakers, and an honest article has to say so. But the mechanism is the same one, the documented cases are specific and serious, and the one study that tested the question directly found that writing from likely-autistic authors was flagged significantly more often.

If you are neurodivergent and a detector has flagged your work, nothing about that flag proves anything about you. Start with the false-accusation playbook; this page gives you the evidence to cite.

The documented cases

The primary source is a Bloomberg Businessweek investigation by Jackie Davalos and Leon Yin (18 October 2024), logged in the AI Incident Database as Incident 849. Businessweek didn't just collect anecdotes — it ran its own test: GPTZero and Copyleaks against "a random sample of 500 college application essays submitted to Texas A&M University in the summer of 2022," before ChatGPT's release, so guaranteed human-written. The result: the services "falsely flagged 1% to 2% of the essays as likely written by AI, in some cases claiming to have near 100% certainty."

The named cases:

• Moira Olmsted, a student at Central Methodist University, received a zero on a reading summary she wrote herself after a detector flagged it. She disputed the accusation "stressing that she has autism spectrum disorder and writes in a formulaic manner that might be mistakenly seen as AI-generated, according to emails viewed by Bloomberg Businessweek." The grade was eventually changed — along with a warning that the next flag would be treated the same way as plagiarism.
• Marley Stevens, a student at the University of North Georgia, was put "on academic probation for a year after a disciplinary hearing determined she'd cheated" over a Turnitin flag; she insisted she wrote the essay herself "using only Grammarly's standard spell-checking and grammar features." (Her case, and what it says about grammar tools, is covered in is Grammarly cheating?.)
• Ken Sahib, "a multilingual student who spent most of his childhood in Italy," received a zero at Berkeley College in New York. His teacher's reply, from emails viewed by Businessweek: "Every tool I tried produced the same result: those responses were AI-generated."

Bloomberg then named exactly the population at risk: "The students most susceptible to inaccurate accusations are likely those who write in a more generic manner, either because they're neurodivergent like Olmsted, speak English as a second language (ESL) or simply learned to use more straightforward vocabulary and a mechanical style."

One to two percent sounds small until you multiply it across a university — the same arithmetic that led Vanderbilt to disable Turnitin's detector: hundreds of wrongly flagged papers a year, with no way to know which ones.

The first empirical study

For years the neurodivergent-bias claim rested on cases like the above. The first study to test it directly is Chambers & Kelley, "The Misclassification of Autistic Writing as AI-Generated" (arXiv:2607.14729). The design: roughly 60,000 Reddit posts split into "likely-autistic" and "general-Reddit" subcorpora, scored with the OpenAI GPT-2 detection model.

The findings, stated carefully, in the paper's own words:

• The bias is there: "significantly more texts from the likely-autistic subcorpus were flagged" as AI-generated.
• The absolute rates were low in this setup — under 2% for both groups.
• The mechanism is not simple: "Connections between features of text with likely-autistic authors and AI-generated text were not straightforward."
• The authors' conclusion: "The widespread use of AI-detection models with a potential bias against autistic writers in their output prompts ethical scrutiny."

Read both halves. This is evidence of a real, statistically significant disparity — and it is one study, using an older detection model and proxy labels (subreddit membership, not diagnoses). It establishes that the concern is empirically grounded, not that the disparity is as large or as well-mapped as the ESL case.

Why the mechanism predicts exactly this

The reason researchers expected this result is that it falls straight out of how detectors work. Detectors score predictability: text with consistent register, regular structure, precise repeated vocabulary, and low variance between sentences reads as "machine-like," because that is the statistical profile of model output.

It is also, frequently, the profile of autistic writing — and of careful writing by people with ADHD using heavy structure to manage drafting, and of anyone taught to write to a formula. The Stanford study that measured the ESL version of this bias (Liang et al.) identified the mechanism precisely: detectors penalize low-perplexity prose, and "non-native English writing samples" were misclassified at an average false-positive rate of 61.3% while native samples sailed through. James Zou, the Stanford study's senior author, put the mechanism in one sentence for Businessweek: "If the word choices tend to be more generic and formulaic, that work has a higher chance of being flagged by AI detectors." The detector does not know why prose is uniform. It punishes uniformity itself — whatever its source.

Even OpenAI, explaining why it shelved its own text-watermarking system, named the fairness problem: detection-style checking could "stigmatize use of AI as a useful writing tool for non-native English speakers." The same logic extends to every writer whose natural register happens to sit near the machine profile.

What the evidence does not yet establish

Honesty requires the other column:

• There is no large-scale measured false-positive rate for neurodivergent writers equivalent to Liang's 61.3% for ESL essays. Nobody has run seven commercial detectors over a diagnosed-writer corpus.
• The Chambers & Kelley study used a GPT-2-era detector and proxy labels; modern commercial detectors may behave differently — in either direction — and no vendor publishes error rates broken down by neurodivergence. Turnitin told Businessweek it "doesn't train specifically on neurodivergent student data or have access to medical histories to assess that classification" — which means the vendor itself cannot rule the bias in or out.
• Case reports, however compelling, cannot establish rates.

None of this weakens the practical conclusion. A detector score was already not valid evidence of misconduct on its own — for anyone. What the neurodivergence evidence adds is a documented reason the error bar is not distributed evenly, and a population with formal disability protections sitting inside it.

If this is you

1. Don't confess to end it. A flag is a statistic, not a finding — by the vendors' own documentation.
2. Use the process playbook: falsely accused of using AI and how to appeal an AI cheating accusation.
3. Name the pattern, with sources. Citing Liang et al., the Bloomberg cases, and Chambers & Kelley reframes the meeting from your honesty to the tool's documented behavior on writers like you. If you have a diagnosis and are comfortable disclosing it, your disability office belongs in the conversation — detector-driven discipline that lands disproportionately on a protected group is an accommodations issue, not just an integrity one.
4. Bring process evidence. Drafts, version history, notes — every type, ranked.

What Olmsted did after her accusation is worth quoting in full, because she independently invented the fix. Per Businessweek: "She screen-recorded herself on her laptop doing writing assignments. She worked in Google Docs to track her changes and create a digital paper trail. She even tried to tweak her vocabulary and syntax." Her reason: "I have so much to lose."

Notice what she did not do: she did not stop being autistic, and she could not stop writing in her own register — tweaking vocabulary to dodge a detector means writing worse on purpose. What actually protected her was the paper trail. If you write in a register detectors keep mistaking for a machine, the durable fix is that same idea, made rigorous: a writing-provenance record kept while you write, sealed against tampering, and checkable by anyone without an account. Your prose style stops being the evidence — the record is. It is free for individual writers, because the people detectors mis-flag are precisely the ones who shouldn't have to pay to be believed.

Go deeper — the primary literature, read closely: non-native speakers and detector bias · how accurate are AI text detectors?