AI, Accessibility & the Neurodivergent Writer

In April 2026 I went through a full neuropsychological assessment at the Brainsight Clinic. Verbal reasoning 86th percentile. Processing speed 23rd. About a twenty-point gap between what I can think and how fast I can get it out of my head and onto a page. The clinical name for that is a processing-speed deficit, and it sits alongside the dyslexia diagnosis I’ve carried since 1983.

I wrote about why this matters for the AI debate in Nobody Knows What They’re Banning. The short version: most arguments about banning AI are arguments about category one, large language models generating creative work. Most accessibility AI use sits in category two, assistive scaffolding. Policy that conflates them lands on the wrong target.

This post is what I didn’t do in that one. The receipts. The clinical literature on what ADHD and dyslexia actually are. The legal framework that already covers what I’m doing. The published research on whether scaffolding tools help or harm cognitive function. The strongest counter-evidence, taken on its own terms. The policy proposal I’d put forward instead of a blanket ban.

If you’ve read the taxonomy argument and want the evidence underneath it, this is the post for you.

Russell Barkley’s 1997 paper “Behavioral Inhibition, Sustained Attention, and the Executive Functions: Constructing a Unifying Theory of ADHD” in Psychological Bulletin 121(1) has over 7,400 citations. It’s the foundational clinical text. His argument: ADHD is a deficit in behavioural inhibition that knocks on into working memory, self-regulation, and reconstitution. His clinical fact sheet “The Important Role of Executive Functioning and Self-Regulation in ADHD” spells out the prosthetic model in his own words: tools that carry execution become for people with executive-function deficits what prosthetic devices are to the physically disabled.

The single most important sentence for the argument I’m making is in there too. Barkley calls ADHD “a disorder mainly of performance rather than of knowledge or skills.” That sentence is doing a lot of work. It means remediation and “trying harder” are the wrong interventions. What works is environmental scaffolding that carries execution forward.

Thomas Brown’s six-cluster model of executive function, in International Journal of Disability, Development and Education 53(1) in 2006, puts processing speed inside the “Effort” cluster and explicitly connects it to difficulty with sustained expository writing. My 86th/23rd split isn’t an anomaly. Braaten et al. (2020, Research in Developmental Disabilities) define a 15-point VCI-PSI gap as a clinically meaningful processing-speed weakness. Mine is about 20. Other studies confirm it as a recognised neuropsych signature of adult ADHD, including Park et al. (2024, Psychiatry Investigation, N=105) and Mikami et al. (2022, Brain and Behavior, N=418 adult ADHD).

Molitor et al., in “Writing Abilities Longitudinally Predict Academic Outcomes of Adolescents with ADHD” (School Psychology Quarterly, 2016, 31(3): 393-404), give the empirical core of what this looks like in practice. People with ADHD produce less organised, shorter, less coherent written text despite equivalent knowledge of the rules. Durand, Arbone and Wharton’s 2020 PeerJ study (N=774) confirmed that adults with ADHD score lower on every organisational-skills measure except strategy knowledge. They know what good writing looks like. What breaks is the execution. Which is exactly why tools that carry execution are the clinically appropriate response.

Sally Shaywitz’s Overcoming Dyslexia (2nd ed., Knopf, 2020) describes the dyslexic profile as a phonological and decoding weakness surrounded by a sea of strengths in reasoning, problem-solving, critical thinking, empathy, and vocabulary. My auditory working memory at the 73rd percentile sits squarely in that sea. Moojen et al.’s 2020 study in Annals of Dyslexia formalises this as the compensatory-skills model: dyslexic adults get to their outcomes through alternative pathways, not despite their profile but in structural relationship with it.

Brock and Fernette Eide’s MIND framework (The Dyslexic Advantage, revised 2023) is useful here. Material, Interconnected, Narrative, Dynamic reasoning. The procedural and rote-memory weakness, which makes orthographic encoding and silent proofreading unreliable for dyslexic writers, isn’t the whole picture. It sits next to genuine narrative and connective-reasoning strengths.

Wood, Moxley, Tighe and Wagner’s 2018 meta-analysis, “Does Use of Text-to-Speech and Related Read-Aloud Tools Improve Reading Comprehension for Students with Reading Disabilities?” (Journal of Learning Disabilities 51(1)), covers 22 studies and 2,942 participants. Weighted effect size 0.35. Statistically significant, peer-reviewed, replicated. My listen-first method, reading everything back through NaturalReader before editing, isn’t a quirk. It’s the documented high-efficacy pathway for my profile.

The Octavia Butler case is worth a moment. Sami Schalk’s 2017 paper “Experience, Research, and Writing: Octavia E. Butler as an Author of Disability Literature,” in Palimpsest: A Journal on Women, Gender, and the Black International, draws on direct archival work with Butler’s papers at the Huntington Library. Butler’s journals carry self-blaming language around periods when her focus failed. Her teachers read her as lazy. Schalk documents her compensation toolkit: she listened to textbooks on tape, subvocalised as she read, drilled herself through repetition and pre-dawn writing discipline. This is also in Lynell George’s A Handful of Earth, A Handful of Sky (Angel City Press, 2020), built from 9,062 pieces in the Butler archive.

Butler was never formally diagnosed with dyslexia in her lifetime. Schalk’s own paper asserts undiagnosed dyslexia directly, drawing on the Butler archive and the Yale Center for Dyslexia and Creativity’s clinical profiling. That caveat about the missing formal diagnosis matters and I want to state it plainly. But it doesn’t weaken the point, because what isn’t attributed is documented directly: her compensation strategy. Books on tape. Subvocalisation. Audio-channel substitution. That was 20th-century TTS. What I’m doing is 21st-century TTS, with an extra layer of scaffolding for the executive-function piece.

John Irving writes by hand because it slows him down. Philip Schultz, diagnosed at 58, writes through endless revision and hundreds of drafts. Dav Pilkey, ADHD and dyslexia both, has said his neurodivergence helped him write stories that weren’t boring. Different tools. Same principle: bespoke scaffolding around weakness, leveraging strength.

The US Department of Education Office for Civil Rights published “Avoiding the Discriminatory Use of Artificial Intelligence” in November 2024. It’s the clearest existing federal authority on this question. It warns explicitly that overly broad bans on AI use can result in disability discrimination under ADA Title II and Section 504, and it obligates institutions to provide reasonable accommodations. Guidance, not binding regulation; enforcement under the current administration is uncertain, but the underlying legal principle predates the guidance.

Tudor v. Whitehall Central School District, 132 F.4th 242 (2d Cir. March 2025), clarifies that an ADA accommodation only needs to be reasonable, not strictly necessary. I don’t need to prove I can’t write without AI. I only need to show that AI is a reasonable accommodation for a documented disability.

The UK Equality Act 2010 (ss. 20-21, Schedule 13) imposes an anticipatory duty to make reasonable adjustments for dyslexia, ADHD, and GAD. The British Dyslexia Association’s “Reasonable Adjustments” guidance names assistive and speech-to-text software as standard adjustments. Disability Rights UK’s guide “Understanding the Equality Act” covers the same ground.

Bezyak, Chesley and Lister’s 2023 report “ChatGPT and Disability: Benefits, Concerns, and Future Potential,” funded by the Rocky Mountain ADA Center and NIDILRR, is the closest thing to an institutional disability-network position. It says blanket ChatGPT bans will hinder the development of increasingly inclusive environments, and it specifically names ADHD, dyslexia, and anxiety disorders as the conditions most clearly benefiting from AI use for idea development, proofreading, organisation, and grammar support.

That list is my exact combination of diagnoses.

The MIT Media Lab preprint from Kosmyna, Hauptmann, Yuan et al. (arXiv:2506.08872, 2025) is the strongest empirical piece of evidence on the other side. EEG study, 54 participants across three sessions, 18 of them returning for a fourth. LLM users showed the lowest neural connectivity, weakest recall of their own text, and a diminished sense of ownership over what they’d written. Effects persisted after the tool was removed. The phrase they use is “cognitive debt.”

This is a real finding and it deserves a real response, not dismissal.

The response is that Kosmyna is a warning about substitution, not scaffolding. The study looked at cognitively intact users using LLMs to write essays they would otherwise have written themselves. My use pattern is structurally different. I’m not outsourcing the thinking. The ideas, structure, emotional architecture, and thematic content exist in documented form before the AI touches any of it. What the AI handles is the surface conversion from notes and intention to grammatically coherent prose, which is exactly the process that a 23rd-percentile processing speed makes unreliable.

Barkley’s distinction, performance not knowledge, maps directly onto this. If your underlying cognitive function is intact and you offload the writing to a machine, you lose the practice that builds the skill. If your underlying cognitive function already has a specific performance deficit that stops the skill being exercised whether a machine is present or not, the calculus is different.

Wu et al.’s 2025 paper in Scientific Reports (N=3,562 across four experimental studies), finding that generative AI collaboration enhances task performance but undermines intrinsic motivation on subsequent solo tasks, belongs here too. Real concern. Same response: the prosthetic model survives it, because the solo task was already constrained by the deficit. There’s no uninstrumented baseline to return to.

Olivia Guest, Iris van Rooij and colleagues, in their 2025 position paper “Against the Uncritical Adoption of ‘AI’ Technologies in Academia” (Zenodo, doi:10.5281/zenodo.17065099), call for universities to treat AI use like essay mills. The paper contains no accessibility carve-out. That omission is documented, and it’s a problem, because the paper is otherwise making a legitimate argument about cognitive skill development and institutional integrity that simply does not reckon with the disability dimension.

I don’t speak for all neurodivergent people. Some of my critics on Bluesky are themselves neurodivergent and have arrived at different conclusions. That’s legitimate, and I want to acknowledge it directly rather than pretend it doesn’t exist. What I’m claiming isn’t that AI use is right for every neurodivergent writer. It’s that it’s defensible for me, given my specific diagnosed profile, the specific functional use I’m making of the tool, and the research base that supports that use.

The concern underneath most of the serious objections isn’t really about AI. It’s about verification. How does anyone know that the ideas, the world, the book, are actually yours?

That’s answerable. I have concept documents predating any AI involvement. I have canonical reference files, world-building bibles, character architecture notes, all built before the first AI session. I have version-controlled manuscript files with a documented audit trail. I have chat logs. I have a full record of what the AI touched and what it didn’t.

This is what good policy looks like. Not a blanket ban. A disclosure and provenance framework. Nature’s 2023 editorial policy is the clearest existing institutional model. AI-assisted copy editing for readability, style, grammar, spelling, punctuation, and tone doesn’t need to be declared. Content generation does. The distinction tracks the functional taxonomy exactly.

The National Science Foundation (December 2023) and National Endowment for the Humanities (October 2024) both run disclosure-based models. AI use permitted with acknowledgement. The NIH NOT-OD-25-132 (July 2025) model, which rejects applications substantially developed by AI with no disability exception, is the hostile-to-accommodation end of the spectrum, and it’s already drawing criticism for exactly that gap.

The proposal I’d put forward is straightforward. Ban content generation. Require disclosure of tool type and function. Recognise accommodation use with a standing accessibility carve-out. Permit provenance audit-trail verification in disputed cases. That isn’t a capitulation. It’s a policy framework that’s actually compatible with disability law.

I’ve been writing this book for years. The ideas are mine. The world is mine. The characters are mine. The spine of the story is mine.

What I use AI for is the surface. The translation of existing thought into coherent prose, a process my processing-speed deficit makes unreliable without scaffolding. This isn’t a shortcut. The AI doesn’t know what happens next. It doesn’t know why a particular scene matters. It doesn’t know what the book is about.

I do.

I was diagnosed dyslexic in 1983. I’ve been finding workarounds for four decades. This is just the most recent one.

Carl Freeman is writing The Faculty of Matter, a dreampunk novel. He writes about the process, the tools, and the long haul at carlmfreeman.substack.com.

Edited with AI