I Was Diagnosed Dyslexic in 1983. AI Is Just the Latest Workaround.

I was diagnosed dyslexic in 1983. I was thirteen.

Thirteen is not a good age to be told you're wired wrong. You've just started secondary school, you're already watching everyone else to figure out what normal looks like, and then someone hands you a clinical label and nobody bothers to explain what it means. They just said the word and expected me to get on with it.

The ADHD diagnosis came much later. April 2026, Brainsight Clinic, full neuropsych assessment. Verbal reasoning 86th percentile. Processing speed 23rd.

That gap matters.

It means there is a measurable distance between what I can think and how fast I can get it out of my head onto a page. That is not a character flaw. It is not laziness. It is not me needing to try harder. It has a name and a clinical framework. The people around me had been saying ADHD for years. The assessment just made it official.

I am also, since 2025, using AI as a writing tool. I have said so on Bluesky and here. Some of the pushback has been worth listening to. Some of it has not. A few people, mostly themselves neurodivergent, have pushed back hard, and that is the response that deserves a careful answer.

This is that answer.

I am not going to pretend AI use is ethically uncomplicated. It isn't. What I am going to argue is narrower. For a writer with a documented neurodivergent profile, using AI as a scaffold around a specific cognitive bottleneck, the research backs what I am doing. The policy frameworks that want to ban it wholesale have not thought hard enough about disability law, accessibility, or the clinical evidence.

People are collapsing unlike things

Not all AI use is the same thing.

That sounds obvious. It gets missed constantly.

Ben Shneiderman's 2020 paper "Human-Centered Artificial Intelligence: Three Fresh Ideas" and his 2022 book Human-Centered AI do the clearest work on this. He separates AI that supports human agency from AI that tries to replace it. Words like amplify, augment and enhance are doing categorically different work from generate and substitute.

Jisc, which is the UK's national centre for digital education, spent most of 2024 and 2025 working through exactly this distinction in their guidance on AI, assessment and accessibility. Same product can be used in fundamentally different ways. Same prompt box generates a story you did not write, or helps you reorganise an outline that is entirely yours.

When people say "AI writing," they usually mean substitution. A machine filling a blank page. That is not what I am doing.

What I am doing is different in kind. The ideas, world, characters and emotional architecture of the book I have been writing for years already exist in notes, canon files, reference material and earlier drafts that predate any AI involvement. What the AI helps with is conversion. It helps me turn existing thought into coherent prose, because coherent prose is exactly where my cognitive profile breaks down.

ADHD is not a knowledge problem

Russell Barkley's 1997 paper "Behavioral Inhibition, Sustained Attention, and the Executive Functions" is the foundational clinical text on adult ADHD. Over 7,400 citations. His later fact sheet on executive function and self-regulation spells it out in plain clinical language.

His argument is that ADHD is not primarily a deficit of intelligence or knowledge. It is a performance problem. He says it directly: ADHD is "a disorder mainly of performance rather than of knowledge or skills."

That sentence does a lot of work.

It means the usual moral language around effort is wrong. Trying harder is not the intervention. Scaffolding is. Barkley's own metaphor is prosthetic: tools that carry execution for people with executive-function deficits do the same work as prosthetic devices for physical disabilities.

Thomas Brown's six-cluster model of executive function (2006) places processing speed inside the Effort cluster and connects it directly to sustained expository writing. That is not abstract for me. My own profile shows the split clearly. Braaten et al. treat a 15-point VCI-PSI gap as a clinically meaningful processing-speed weakness. Mine is 20. Park et al. (2024, N=105) and Mikami et al. (2022, N=418) confirm it as a recognised adult ADHD signature.

The practical consequence is simple. People with ADHD can know exactly what good writing looks like and still not produce it in organised, sustained form. Molitor et al. showed this in adolescents. Durand, Arbone and Wharton showed it in adults. The rule knowledge is there. The strategy knowledge is there. The execution is what breaks.

Which is why a tool that carries execution is not cheating the underlying capacity. It is compensating for the bottleneck.

Dyslexia is not just a deficit story

The same is true of dyslexia.

Sally Shaywitz's Overcoming Dyslexia describes the profile as phonological and decoding weakness sitting next to major strengths in reasoning, problem-solving, empathy, critical thinking and vocabulary. Moojen et al. formalise this as the compensatory-skills model. Dyslexic adults do not succeed despite the profile. We succeed by building alternative pathways around it.

Brock and Fernette Eide's MIND framework, from The Dyslexic Advantage, is useful shorthand here. Material, Interconnected, Narrative, Dynamic reasoning. It names something dyslexic people usually know before clinicians describe it: the difficulty is real, and it sits next to real strengths in pattern, connection and synthesis.

This is why my use of text-to-speech is not incidental. Wood, Moxley, Tighe and Wagner's 2018 meta-analysis on read-aloud tools, covering 22 studies and 2,942 participants, found a significant positive effect. Effect size 0.35. Peer-reviewed, replicated, uncontroversial. My listen-first method is not a quirk. It is the documented high-efficacy pathway for my profile.

There is a historical example here worth taking seriously.

Octavia Butler was never formally diagnosed with dyslexia. That caveat matters and I want to state it plainly. But Sami Schalk's 2017 archival work at the Huntington Library shows a writer who built a serious compensation toolkit around exactly the difficulties many dyslexic writers recognise. Books on tape. Subvocalisation. Relentless repetition. Pre-dawn writing discipline. Lynell George's work on the same archive, 9,062 pieces of it, shows the same pattern.

The formal label is missing. The compensation method is not.

What Butler did with tape and repetition, I do with TTS and other scaffolds. Different century. Same logic.

John Irving writes by hand to slow himself down. Philip Schultz, diagnosed at 58, writes through hundreds of drafts. Dav Pilkey has spoken openly about how ADHD and dyslexia shaped the stories he writes. Different writers. Different tools. Same principle: build a method around the weakness without surrendering the work.

The legal question is not optional

A lot of the AI discourse has not caught up with this part.

The US Department of Education's Office for Civil Rights published guidance in November 2024 warning that overly broad AI bans can create disability discrimination problems under ADA Title II and Section 504. It is guidance, not statute. Its enforcement future is uncertain. The principle underneath it is not new.

The more important point is conceptual. If a tool is being used as a reasonable accommodation for a documented disability, blanket prohibition becomes much harder to defend.

Tudor v. Whitehall Central School District (2025) clarifies that an accommodation does not need to be strictly necessary to be reasonable. I do not have to prove I am literally incapable of writing without AI. I have to show it is a reasonable accommodation for a documented impairment.

The UK framework points the same way. Equality Act 2010, ss. 20-21, Schedule 13. Anticipatory duty to make reasonable adjustments covering dyslexia, ADHD and GAD. British Dyslexia Association guidance names assistive and speech-to-text software as standard adjustments.

The most directly relevant disability-sector document I found was Bezyak, Chesley and Lister's 2023 report on ChatGPT and disability, funded by the Rocky Mountain ADA Center. They are not subtle about it. Blanket bans risk blocking more inclusive environments. They name ADHD, dyslexia and anxiety disorders as the conditions most clearly benefiting from AI use for organisation, proofreading, grammar and idea development.

That is, inconveniently, my exact combination.

The objections are real, and I take them seriously

There is a serious case on the other side and it should be faced honestly.

The strongest recent evidence is the MIT Media Lab preprint by Kosmyna, Hauptmann, Yuan and colleagues. EEG study. LLM users showed weaker neural connectivity, worse recall of their own text, and a diminished sense of authorship. Effects persisted after the tool was removed. They call it cognitive debt.

That is not nothing. It is a real warning.

But it is a warning about substitution, not about every possible form of scaffolding.

The participants in that study were cognitively intact users using LLMs to produce essays they would otherwise have written themselves. My use case is structurally different. I am not outsourcing the ideas, narrative logic, world design or emotional architecture. Those are already there. What I am scaffolding is the bottlenecked conversion of existing thought into coherent written form.

That is a different thing.

Barkley comes back here. If your underlying performance system is intact and you hand the whole task to a machine, you may well be eroding a skill you should be practising. If your underlying problem is a performance deficit that blocks execution even when the ideas are present, the calculus changes.

Wu, Liu, Ruan, Chen and Xie's 2025 Scientific Reports paper (N=3,562) sits in the same category for me. Real concern. Real finding: AI collaboration enhances performance but reduces intrinsic motivation on subsequent solo tasks. But again, it assumes a stronger unassisted baseline than my profile actually gives me. There is no clean "just do it without the tool" version waiting on the other side. The bottleneck is already there.

The Guest and van Rooij position paper against uncritical AI adoption in academia matters too. I understand the institutional concern. I agree with parts of it. But the paper contains no meaningful accessibility carve-out, and that omission weakens it badly. If your policy model treats all AI use as equivalent to an essay mill, you are not just being strict. You are flattening disability into misconduct.

I also want to say this plainly. I do not speak for all neurodivergent people.

Some of the strongest criticism I have had has come from neurodivergent people themselves. That is legitimate. I am not arguing AI is right for every dyslexic or ADHD writer. I am arguing something narrower. It is defensible for me, given my diagnosed profile, my actual use pattern, and the research that maps onto it.

The real issue underneath this is provenance

Most serious objections eventually land here.

The question is not really "AI or no AI." The question is: how does anyone know the underlying work is yours?

That is answerable.

I have concept documents that predate AI involvement. I have canon files, world bibles, architectural notes, character work, version-controlled manuscript records. I have chat logs. I have an audit trail. I can show what existed before AI touched anything and what AI actually did afterward.

That is where sensible policy should go. Not a blanket ban. Not a panicked shrug. A provenance framework.

Nature's 2023 editorial policy is still the clearest institutional example I have seen. AI-assisted copy editing for readability, grammar, spelling, punctuation and tone does not require declaration. Content generation does. That distinction tracks the actual functional difference far better than most university policy does.

The NSF (2023) and NEH (2024) both run disclosure-based models. The NIH's 2025 model sits at the much more hostile end, rejecting substantially AI-developed applications without any disability carve-out. That is exactly the kind of framework that is going to attract legal and ethical criticism, and deservedly so.

The policy answer seems straightforward. Ban content generation. Require disclosure of tool type and function. Recognise accommodation use with a standing carve-out. Preserve provenance audit trails for disputed cases.

That is not capitulation. It is what sane policy looks like if you want intellectual integrity without trampling disability law.

Where I stand

I have been writing this book for years.

The ideas are mine. The world is mine. The characters are mine. The emotional spine is mine.

What AI helps me with is surface conversion. It helps me move existing thought into coherent prose in a way my processing-speed deficit does not reliably allow on its own.

That is not a shortcut.

The AI does not know what happens next. It does not know why a scene matters. It does not know what the book is about.

I do.

I was diagnosed dyslexic in 1983. I have been finding workarounds for more than forty years.

This is just the latest one.

Carl Freeman is writing The Faculty of Matter, a dreampunk novel. He writes about process, tools and the long haul on Substack.