Edition 369 | August 6, 2026
The Dyslexic AI Newsletter by LM Lab AI

Reading Time: 8 minutes | Listening Time: 11 minutes

Five editions ago I told you I was going to take the six-part series that started with a Feynman video at 5 AM and turn it into one real essay. Maybe a white paper. I said I'd let you know when it was ready.

It's not ready. I want to own that up front instead of quietly hoping you forgot.

But something landed in the research this week that I think is the first real chapter of that essay, and it didn't come from me. It came from a lab.

Researchers at EPFL, that's École Polytechnique Fédérale de Lausanne, a top technical research university in Switzerland, recently published a study that sounds almost too on-the-nose to be real. They took a vision-language model, found the part of it that functions like a visual word-form area (the region that does the specific work of recognizing written words), and disabled it. Just that part. Everything else about the model's reasoning, its image understanding, its general language ability, stayed intact.

What they got was an AI that reads like a dyslexic brain reads, while thinking just as clearly as it did before.

One of the researchers, Martin Schrimpf, put it plainly: you can't ethically walk into a human brain and knock out a specific set of neurons to see what happens. In a model, you can. That's the actual breakthrough here. Not that dyslexia got "solved," but that for the first time there's a computational framework for studying what's actually happening, mechanistically, instead of only through self-report and behavioral testing.

Once they had a model that struggled with reading the way a dyslexic person does, they didn't stop there. They started testing fonts against it.

Dyslexia-friendly typefaces have been a topic of debate for years. Plenty of anecdote, mixed research, a lot of "this helped me, your mileage may vary." What EPFL's disabled model gave them was something closer to an objective instrument. Feed it a font built with dyslexic readers in mind, and it performed measurably better. Feed it a font known to be harder to parse, and it got measurably worse. Same underlying model, same disabled region, different result, purely based on the shape of the letters.

That's not a vibe anymore. That's a mechanism you can point to.

Go back to Edition 367, the Wave Model of Meaning. The whole argument there was that meaning in these systems isn't a fixed dictionary entry. It's relational, a field that shifts depending on context, and that a lot of dyslexic minds already operate that way natively.

Here's what the EPFL work adds to that. When you strip away the specific symbol-recognition shortcut and leave everything else standing, the system doesn't get dumber. It gets a different relationship to the page. What changes is exactly the narrow thing this newsletter has been pointing at since Edition 362. Not the thinking, the tax on proving the thinking through one specific channel.

That's the whole thesis, sitting inside a mechanistic AI study that has nothing to do with education policy or newsletter writing. Nobody there was trying to prove me right. They were just trying to understand a brain difference well enough to design better fonts. And the byproduct is about as clean a piece of outside validation as this argument is ever going to get.

Here's the second piece, and it's the practical half.

A separate, newly published framework, researchers are calling it the "AI Scaffold and Engagement Spectrum" (you can read it here), lays out how to actually build AI support around students with dyslexia without either ignoring the tool or letting it do the thinking for them. One layer sets clear, assignment-specific rules for where AI help is appropriate. The other gives the student an actual structured workflow: text-to-speech, visual organizers, chunked instructions, for reading, planning, drafting, and revising with AI alongside them, while still doing the real cognitive work themselves.

That is, almost word for word, the thing I was circling in Edition 368 with the six pillars and the AI-Native Microcollege idea. I was describing it from the outside, as a parent watching what worked for Makena. This is researchers building the actual scaffolding, with explicit rules instead of vibes.

Put the two studies next to each other and you get the full shape: one explains why a dyslexic mind meets friction in a system built for a different kind of processing. The other shows how to build real support around that difference instead of either ignoring it or outsourcing the thinking entirely.

I'm not going to pretend this counts as the finished essay. It doesn't. But I wanted you to see the first real external data point landing, because it's a much better foundation than six weekly newsletter editions and my own conviction. When I do sit down to write the full piece, this is going in it. First citation, not last thought.

One thing to try this week.

If you or your kid reads on a screen regularly, actually test a dyslexia-friendly font instead of assuming the default one is fine. OpenDyslexic and Lexend are both free and easy to try. Five minutes, side by side, see if anything shifts. You don't need a lab to run this experiment. The EPFL team just gave you a reason to actually believe it might matter.

Use this to build your own version of the "Engagement Spectrum" for something you or your kid is working on right now.

"Here's a task I (or my kid) am working on: [describe it]. Help me build an 'engagement spectrum' for it: specific rules for which parts I should do completely on my own, which parts you can help scaffold (like organizing, reading aloud, or breaking into steps), and which parts would cross the line into you doing the actual thinking for me. Be specific to this exact task, not generic."

Know where the line is on purpose. Then use everything on your side of it, fully.

Matt "Coach" Ivey

Founder, LM Lab AI | Creator, The Dyslexic AI Newsletter

Dictated, not typed. Obviously.

Still open, but there's a real chance the promised synthesis essay finally gets started soon, now that it has actual outside research to stand on instead of just this newsletter's word for it.