AI Detectors Flag Autistic and ADHD Writers — The Evidence

AI Detectors Flag Autistic and ADHD Writers: The Evidence

The best-documented failure of AI writing detectors is their bias against non-native English speakers—peer-reviewed, quantified, and now widely conceded. A second group is showing up in the same trap, for the same underlying reason, but with far less attention and almost no hard data: neurodivergent students. Autistic and ADHD writers are being flagged as "AI," and the cases are already reaching courtrooms. Here's what the evidence does and doesn't show.

Why neurodivergent writing reads as "AI"

AI detectors work by measuring how statistically "surprising" your writing is—low surprise (predictable word choices, even rhythm, consistent structure) looks like a machine; high surprise looks human. Researchers call these signals perplexity and "burstiness."

The problem is that a lot of authentic neurodivergent writing is, by its nature, low-surprise. University teaching centers have started warning faculty about exactly this. The University of Nebraska–Lincoln's teaching center notes that neurodivergent students—"autism, ADHD, dyslexia"—"are also prone to receive false positive ratings," because traits like leaning on repeated phrases and a consistent, formulaic structure get "misconstrued as AI writing." Northern Illinois University's teaching center likewise flags that "neurodiverse students are also more likely to be falsely flagged."

The mechanism is the same one that produces the documented non-native-speaker bias: writing with steady rhythm and a constrained, repeated vocabulary sits in the statistical overlap between human and machine text. An autistic student's precise, literal, meticulously structured prose and an ADHD student's focused, repetitive bursts both land in that overlap. This is why it isn't a tuning bug a better detector fixes—it's the same structural floor we explain in why AI-detector false positives are mathematically unavoidable.

The cases that put faces on it

Three cases show how this plays out—two autistic students and one with documented anxiety and OCD:

• Moira Olmsted (Central Methodist University, 2023). An autistic student studying to become a teacher had a weekly assignment scored zero after a detector flagged it; her "naturally formulaic writing style," shaped by being on the spectrum, was misread as machine output. She had to fight to clear her name. Her story anchored a Bloomberg investigation into detector false accusations.
• Orion Newby (Adelphi University, decided January 2026). A freshman on the autism spectrum, enrolled in Adelphi's neurodevelopmental support program, was accused after a professor cited a Turnitin AI score of "100%." Newby produced two other detectors (Grammarly and ZeroGPT) that called the essay human-written. A New York state court annulled the finding and ordered his record expunged, calling it "without valid basis." It's one of the cases in our AI cheating lawsuits tracker.
• "Jane Doe" v. University of Michigan (filed 2026). A student with generalized anxiety and OCD sued in federal court, alleging her disability-related writing traits—"formal tone, meticulous structure, stylistic consistency"—were treated as evidence of AI use, and that the instructor reached that conclusion through a "circular methodology." Her claim is framed explicitly under the Americans with Disabilities Act and the Rehabilitation Act.

That last case points at the real legal exposure. When a school disciplines a student based on a detector score that fired because of a disability-linked writing style, it isn't just an academic-integrity dispute—it's a potential disability-discrimination problem.

The disability-rights frame

Accusing a student on a detector alone risks running into Section 504 of the Rehabilitation Act and the ADA, which bar treating disability-related characteristics as misconduct and require an accessible, fair process. Disability-rights researchers saw this coming: the Center for Democracy & Technology documented years ago how automated "flagging" technologies in education systematically disadvantage disabled people—surveillance built around a default, "normal" user, where simply being different enough trips the alarm. Advocate Lydia X. Z. Brown's framing fits AI detection exactly: disabled people end up "simultaneously hyper-surveilled" and presumed suspicious by tools that were never designed with them in mind.

Notably, the warnings are coming from teaching-and-learning centers, not yet from a wave of formal university disability-office policies—which is part of the problem. The institutions best positioned to protect these students haven't caught up to the tools being used against them.

What this means—and what it doesn't

Let's be precise, because overclaiming here would be its own kind of harm:

• What's solid: the mechanism (low-perplexity writing reads as AI), the peer-reviewed parallel (the 61% non-native-speaker false-positive rate, driven by the identical signal), and named cases of autistic and disabled students wrongly flagged, one already overturned in court.
• What's missing: a controlled study measuring how often neurodivergent writers are falsely flagged. That gap is exactly why this is the emerging frontier and not a settled finding—and why circulating "X% of autistic students are flagged" statistics should be treated as invented until a real study exists.

The practical takeaway for a neurodivergent student is the same one the math points to for everyone: a detector score is a guess, not proof. Keep your drafts and version history as evidence of your process; if you're accused, ask for the policy and a human review, and—if relevant—loop in your disability-services office, because a process that can't accommodate how you write may be the institution's problem, not yours. For more on the broader pattern of detectors misfiring on the people least able to absorb it, see our work on false-positive rates and the stories of students falsely accused.

---

Sources

University of Nebraska–Lincoln Center for Transformative Teaching ("The Challenge of AI Checkers"); Northern Illinois University CITL ("AI detectors: an ethical minefield," 2024); Liang et al., "GPT detectors are biased against non-native English writers" (Patterns, 2023) for the shared mechanism; Bloomberg, "Do AI Detectors Work? Students Face False Cheating Accusations" (2024) on Moira Olmsted; reporting on Newby v. Adelphi University (Nassau County Supreme Court, 2026); reporting on the University of Michigan disability-discrimination suit (2026); Center for Democracy & Technology, "Ableism and Disability Discrimination in New Surveillance Technologies" (Brown & Shetty, 2022). No quantified neurodivergent false-positive rate has been published; claims of one should be verified against a primary study. Nothing here is legal or medical advice.

Related Reading

• AI Detector False Positives Are Mathematically Unavoidable
• AI Detection Tools Are Biased Against International Students
• AI Cheating Lawsuits Tracker — Every Case, Who Won
• Flagxiety Stories: Students Falsely Accused by AI Detectors
• AI Detector False-Positive Rates Compared