AI Detectors and Neurodivergent Writers: When Structure Looks Like a Machine

If an AI detector flagged your essay and you are autistic, have ADHD, or are dyslexic, the short version is this: you are almost certainly tripping the same statistical wire that catches non-native English speakers, and it says nothing about whether you used AI. Detectors flag text with low perplexity — text where each next word is statistically predictable from the words before it. Writing that is consistent, carefully structured, and built from reliable phrasing scores as predictable. For many neurodivergent writers, that is not a flaw to correct. It is simply how their writing works.

The second thing to know: you cannot reliably write your way around a detector without damaging your writing, and the better move is to stop trying. Protection comes from evidence of process — version history, drafts, notes, and, if you want something stronger, a cryptographic record of your writing sessions. This post walks through what neurodivergent writers have been reporting, why the mechanism singles out structured writers of every kind, and the documentation habits that end an accusation conversation quickly.

What neurodivergent writers keep reporting

I have not run a study on this, and I will not pretend one exists. What does exist is a steady accumulation of first-person reports — in student forums, in disability communities, in threads where teachers compare notes — and they describe the same event with different details.

An autistic student writes the way they have always written: formal register, precise repeated terminology, sentences that follow a consistent internal template. The detector calls it AI. An ADHD student, who finishes essays only by leaning hard on an outline — first point, evidence, transition, next point — gets flagged for exactly that scaffolding. A dyslexic student runs a draft through a spelling and grammar checker, as they have been told to do since primary school, and the smoothed-out text comes back “likely AI-generated.”

Teachers appear in these same threads, often uneasy. Many say openly that they do not know what to do with a detector score: the tool hands them a percentage and no way to interrogate it. The flag arrives with the authority of software and none of the accountability of a person.

None of this means detectors were built to be biased against neurodivergent people. Nobody engineered that. It is a side effect of what the tools actually measure — and that is worth understanding precisely, because it explains why the advice “just write more naturally” is useless.

The mechanism: detectors punish predictable text

Most AI detectors estimate two things about your text. The first is perplexity: how surprised a language model is by each next word, given the words before it. AI-generated text tends to have low perplexity, because the model literally chose the most probable words. The second is often called burstiness: how much your sentence length and structure vary. Human writing supposedly swings between long and short, complex and plain; machine text is supposedly uniform.

Notice what these metrics describe. Not authorship — style. Specifically, a style with high variance and unpredictable word choice. If your natural writing has low variance and predictable word choice, you look like a machine to the machine.

Now put common neurodivergent writing traits next to those signals:

Every row of that table is a legitimate — often hard-won — writing strategy. Every row also pushes the score in the wrong direction. That is the whole problem: the detector cannot tell disciplined from generated.

This wire has been tripped before

There is one rigorous data point that shows the mechanism does real damage. A Stanford study (Liang et al., published in Patterns, 2023) found AI detectors falsely flagged 61% of TOEFL essays by non-native English speakers. That failure maps onto exactly the dynamics above: writers working in a second language draw on a smaller active vocabulary and rely on learned sentence patterns, which lowers perplexity, which triggers flags.

I write English as a second language, and that study describes my writing life exactly. What the neurodivergent reports suggest — and what the mechanism predicts — is that ESL writers were never a special case. They were simply the first measured population of a much larger group: everyone whose writing is more regular than the detector’s model of “human.” If detectors misfire this consistently on one measurable group, treating their scores as evidence against any individual writer is indefensible — a point covered in more depth in Is Turnitin’s AI detection accurate?.

Build your evidence before anyone asks for it

You cannot control the detector. You can control what happens in the thirty seconds after someone shows you a score. The writers who get through these accusations fastest are the ones who can immediately show process — not the ones with the best argument about false positive rates.

Practical habits, roughly in order of effort:

Write where history accumulates. Google Docs and Word Online keep version history automatically. Write in the document from the first sentence. Do not compose elsewhere and paste in — a paste of 800 words appearing at once looks like exactly the thing you are trying to disprove.

Keep the mess. The outline, the reading notes, the abandoned half-sentences, the comment to yourself that says “MOVE THIS SECTION.” For ADHD writers especially, the mess is the proof: no chatbot output includes four false starts.

Save dated drafts at milestones. Version history can be awkward to present in a meeting; three dated draft files are easy.

Consider a signed process record. This is what Diglot’s Authorship Certificate is for: as you write in the editor, it records the sequence of writing events — typing, pausing, revising — into a cryptographically signed, append-only chain. The result is not a “human score” from another black box. It is a verifiable timeline of how the document came to exist, which you can hand to an instructor without asking them to trust a vendor’s percentage.

Whichever tools you choose, set them up before the semester starts, not after the email arrives.

Talking to instructors without disclosing more than you want

You do not owe anyone a diagnosis to defend your own writing. It helps to separate two conversations that often get tangled together.

The process conversation is available to everyone: “This is my work. Here are my drafts, notes, and version history — I’m happy to walk you through how the essay was built.” This conversation never mentions neurodivergence, and in most cases it is the stronger one, because it answers the actual question — did you write this? — with evidence rather than context.

The accommodation conversation is optional and entirely yours to choose. If you are registered with your institution’s disability services, you can ask whether a note about your writing profile can be attached to your file, and whether the department has a policy on detector use — some departments now decline to treat scores as standalone evidence. If you would rather not disclose, that is a legitimate choice; lean on the process conversation instead.

If a flag has already landed: do not rewrite the essay to “sound more human,” and be careful about accepting a redo “just to be safe” — both can read as quiet admissions. There is a step-by-step guide to appealing a false AI-writing accusation that applies here unchanged.

The quiet cost nobody grades

The reports carry a second layer that gets less attention than the accusations themselves: what the fear does to the writing. Writers describe deliberately roughening their sentences, breaking the structure that lets them finish work at all, un-fixing grammar a tool had corrected — degrading their own writing to seem more plausibly human. There is a name for this fear — flagxiety — and neurodivergent writers may be paying its highest price, because the fix being implicitly demanded of them is to write like someone else.

That is the part worth refusing. The consistent structure that carries an ADHD writer to the final paragraph, the precise repeated terminology an autistic writer chose on purpose, the cleaned-up text a dyslexic writer produced with ordinary tools — these are not tells. They are technique. The failure sits in a measurement that cannot tell technique from automation, and the answer is to document your process, not dismantle it.

I built Diglot because my own writing — second-language English, learned patterns and all — is exactly the kind detectors distrust. If yours is too, the Diglot editor records your writing process into an Authorship Certificate as you work, so the proof exists before anyone asks for it. Write the way that works for you, and let the record answer for itself.