Edition 377 | The Dyslexic AI Newsletter by LM Lab AI

- The podcast that made me question my own assumptions about AI and neurodivergent thinkers

- What happened when I asked AI to argue against my own position

- Why my experience with AI might not generalize to everyone

- The difference between AI substitution, augmentation, and accessibility

- Why generic AI models fail neurodivergent users

- What I'm changing about how Cognitive Partner should work because of this

Reading Time: 8 minutes | Listening Time: 10 minutes

🧠 I've spent a lot of time thinking about how AI can help dyslexic and neurodivergent thinkers. But that doesn't mean I have it all figured out.

In fact, I recently listened to a podcast that made me stop and question some of the things I've been saying.

Not because I suddenly think AI is bad.

I don't.

I've personally experienced too many benefits from these tools to believe that.

But I realized something important.

My experience with AI is also a source of bias.

And if I'm going to spend my time talking about AI and neurodivergent thinking, I need to be willing to question my own assumptions.

I listened to an episode of Hyperfocus with Rae Jacobson from Understood called "What could the AI boom mean for neurodivergent people?"

Dr. Amy Gaeta, an AI ethicist and researcher, talked about both the potential benefits and risks of AI for neurodivergent people.

There were several things I agreed with.

AI can reinforce our existing beliefs. AI can give incorrect information. AI can create problems when we blindly trust the output. AI can potentially replace important learning processes. And AI designed without neurodivergent people in the room can make assumptions about what intelligence, productivity, learning, and accessibility are supposed to look like.

All of those concerns are worth taking seriously.

But one part really stuck with me.

The conversation around AI doing the thinking for us.

That made me ask myself: Am I actually using AI to think for me? Or am I using it to help me think?

This is where it got interesting.

Instead of asking AI to agree with me, I asked it to challenge my position.

What am I missing? Where am I biased? What assumptions am I making because AI has worked so well for me?

And it gave me some pretty good answers.

One of the biggest things I realized is that I may be extrapolating from my own experience.

I'm a dyslexic thinker who has spent a lot of time experimenting with AI. I've built workflows around it. I've experimented with agents. I've created ways for AI to challenge me. I don't just ask a question and blindly accept the answer. I often ask it to argue the opposite side. I ask it to find holes in my thinking. I ask it to give me evidence. I ask it to explain what it doesn't know. I ask it to tell me when I might be wrong.

I've essentially been building what I call a Cognitive Partner.

But here's the problem:

Most people aren't using AI that way.

And even if they were, that doesn't mean every person should.

One of the points made in the podcast was that reading can be a process-oriented task.

For some people, the act of reading and working through a research paper is part of how they learn.

I understand that.

But this is where my experience as a dyslexic thinker changes the equation.

Give me a 20-page academic research paper filled with technical language, unfamiliar terminology, statistics and methodology, and the barrier isn't necessarily my ability to understand the ideas.

The barrier can be getting through the document in the first place.

And there is an enormous amount of research being published every day.

I created a research tool that helps me deal with this. It can summarize a paper based on what I already know, what I'm researching and what I'm actually trying to understand. It can identify the sections that are most relevant to me. It can explain technical concepts in language I can process.

And then I can go back to the original research.

I can ask it: "Show me where that came from." "What did you leave out?" "Which part of this paper should I actually read?" "What evidence supports that conclusion?" "What does the opposing research say?"

That's a very different experience from simply asking AI to read the paper and tell me what to believe.

I'm not necessarily outsourcing the thinking.

I'm removing a barrier to accessing the information.

And I think that's an important distinction.

Accessibility doesn't automatically mean something is good.

Sometimes friction is useful.

A child learning to read needs opportunities to actually develop reading skills. A student learning to write needs opportunities to construct their own ideas. A researcher may need to wrestle with a difficult paper because that process is part of their education.

So we can't simply say: "AI makes this easier, therefore AI should do it."

That's too simplistic.

We need to ask: What is the purpose of the task?

And perhaps more importantly: What are we giving up when we delegate it?

That answer may be completely different for a 10-year-old than it is for a 40-year-old.

That distinction is something I hadn't given enough weight to.

This is one idea I want to explore more.

Maybe we shouldn't judge AI based solely on what task it performs.

We should ask what cognitive function it is replacing.

If AI does all of the reading for a child who is supposed to be learning to read, that's potentially very different from AI helping an adult dyslexic researcher access a 20-page technical paper.

The same technology can have completely different consequences depending on the person, the task and the stage of development.

That's why I'm becoming less interested in asking: "Is AI good or bad for neurodivergent people?"

And more interested in asking: "What kind of AI is good for which person, doing which task, at which point in their development?"

That's a much harder question. It's also a much more interesting one.

This is where I think this conversation changes the way I think about Dyslexic AI.

I don't want a Cognitive Partner that simply agrees with me. I don't want an AI that always makes things easier. And I definitely don't want an AI that quietly makes decisions for me.

I want something that understands when to assist, when to explain, when to challenge, when to ask questions, when to show its sources, and when to tell me that I should probably do the work myself.

That means the system needs to understand context. It needs to understand the individual. It needs to understand the task. And it needs to understand the difference between removing an unnecessary barrier and removing a valuable learning experience.

That's a much more complicated problem. But I think it's the right problem.

Another thing from the conversation that I strongly agree with is the problem with generic AI.

A general-purpose model doesn't necessarily understand the specific needs of someone with dyslexia, ADHD, autism, dyscalculia, dysgraphia or another learning difference. And it certainly doesn't understand every individual's needs.

That's one reason I believe so strongly in specialized AI systems.

Imagine organizations that actually work with these communities helping design the models, datasets, prompts, evaluations, guardrails and workflows. Instead of simply putting a generic chatbot in front of someone and saying: "Good luck."

We could build systems specifically designed to understand the population they're serving.

That doesn't mean the AI will never be wrong. It means we're designing the system to make fewer predictable mistakes and to behave appropriately when it doesn't know something.

The example discussed in the podcast about a chatbot giving inappropriate weight-loss advice to someone seeking eating-disorder support is exactly why this matters.

That's not an argument against AI.

To me, it's an argument for better AI architecture, better guardrails and better human oversight.

I have a tendency to think: If AI helps me do something I couldn't easily do before, that's a good thing.

And sometimes it is.

But I need to remember that I'm not everyone.

What works for me might not work for another dyslexic thinker. What works for an adult may not work for a child. What works for someone who understands how to interrogate an AI may not work for someone who simply accepts the first answer. And what works today may not work six months from now.

That's the strange thing about building anything around AI.

The ground keeps moving.

Technology normally moves much faster than the institutions built around it.

A new textbook can take years. A new study can take years to influence practice. A new educational method can take years to make its way through a school district. A new workplace accommodation can take years to become standard practice.

AI gives us a different possibility.

We can change the instructions. We can change the workflow. We can change the model. We can add new research. We can add guardrails. We can test new approaches. We can adapt the tool as we learn.

We may always be trying to catch up with the technology.

But we can potentially close the gap much faster.

And that might be one of the most exciting parts of all of this.

This is probably the biggest thing I took away from this conversation.

I don't want Dyslexic AI to become a project where I prove that my original ideas were correct.

I want it to become a project where we keep figuring out what works.

That means listening to researchers. Listening to educators. Listening to therapists. Listening to parents. Listening to developers. Listening to neurodivergent people.

And listening to people who disagree with me.

Especially those people.

Because if I'm building a Cognitive Partner that is supposed to challenge people's thinking, I probably need to be willing to let other people challenge mine.

I'm still learning how to use these tools. I'm still learning where they help. I'm still learning where they can hurt. And I'm still learning what a truly useful Cognitive Partner should actually be.

That's not a weakness in the idea.

I think it's part of the idea.

AI is changing too quickly for any of us to have the final answer.

So maybe the goal isn't to build the perfect model.

Maybe the goal is to build systems that can keep learning, keep adapting and keep listening to the people they're designed to help.

And if we can do that, I think there's a real opportunity here.

Not to make neurodivergent people think more like machines.

But to build technology that helps more people think in the ways their brains already work best.

And I'm still figuring out what that looks like.

That's the fun part.

Pick something you believe strongly about AI, about dyslexia, about anything. Ask AI to argue against it, not politely. Ask it where you're biased, what you're missing, and what you're only confident about because it's worked for you personally.

You don't have to change your mind. But you should know if you can defend it.

"Challenge my position on [your topic]. What am I missing? Where am I biased? What assumptions am I making because this has worked well for me personally? Don't soften it, and don't just agree with me because I asked."

- Hyperfocus with Rae Jacobson: "What could the AI boom mean for neurodivergent people?", featuring Dr. Amy Gaeta, AI ethicist and researcher at the University of Cambridge.

Matt "Coach" Ivey

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

Dictated, not typed. Obviously.

- 🎧 A podcast (Hyperfocus, with Dr. Amy Gaeta) made me question some of what I've been saying about AI and neurodivergent thinking.

- 🪞 I asked AI to argue against my own position instead of agreeing with me. It found real gaps.

- 🧍 My experience with AI is also a bias. What works for me doesn't automatically work for every dyslexic thinker, or every age.

- 📚 For me, a dense research paper's barrier is getting through it, not understanding the ideas. AI removing that barrier isn't the same as AI doing my thinking.

- 🔀 Three different things get called "AI use": substitution (it thought for me), augmentation (it helped me think), accessibility (it removed a mechanical barrier).

- 🏗️ Generic AI models don't understand dyslexia, ADHD, autism, or any specific learning difference well. That's the case for building specialized systems, not generic chatbots.

- 🔁 I don't want to be right about Cognitive Partner. I want to keep listening, especially to people who disagree with me.

## Previously

- Edition 376: "When Is Offloading Actually Learning?" (offloading vs. outsourcing vs. augmentation)

- Edition 375: "They're Giving the Bots ID Cards First" (proof of human, agent internet)

- Edition 374: "I'm On AI's Side and I Can't Tell Anymore" (Cognitive Provenance)

Next edition: "Fix the Text, Not the Reader" (CogniLens) — parked for now, still needs a gut-check on the favorite-author placeholder and the Calacanis section before it's ready.