Edition 350 | April 23, 2026 The Dyslexic AI Newsletter by LM Lab AI

Reading Time: 9 minutes Listening Time: 12 minutes

Happy Monday,

May the 4th be with you 😉

A paper dropped out of MIT's CSAIL lab last week that I cannot stop thinking about.

The headline: researchers taught AI to say "I am not sure."

That sounds simple. It is not. It is one of the most important findings in AI training I have seen this year. And it connects to something I have been writing about, in different forms, for longer than I realized.

Let me explain.

Today's reasoning AI models have a confidence problem.

When you ask one of them a question, it answers with the same level of certainty whether it actually knows the answer or it is essentially flipping a coin. The most powerful systems in the world will tell you something with complete confidence and be completely wrong, and you have no signal to know the difference.

MIT researchers traced this back to a specific flaw in how the models are trained.

The standard approach: reward the model for getting the right answer. Penalize it for getting the wrong answer. Nothing in between.

The result: the model learns to confidently produce an answer no matter what. There is no incentive to express uncertainty. There is no reward for saying "I am not sure." The training process literally teaches confident guessing.

In MIT's words, "The standard training approach is simple and powerful, but it gives the model no incentive to express uncertainty or say I do not know. So the model naturally learns to guess when it is unsure."

That is a problem with consequences.

In medicine, finance, law, or any setting where people make real decisions based on AI output, a system that says "I am 95 percent sure" when it is actually right half the time is more dangerous than a system that gets the answer wrong outright. Because there is no warning signal.

You can spot a wrong answer. You cannot easily spot a wrong confidence level.

Here is what they did.

They built a method called RLCR, which stands for Reinforcement Learning with Calibration Rewards. Instead of just training models to be right or wrong, they added a single new term to the reward function: a measure that penalizes the gap between how confident the model says it is and how often it is actually correct.

Now the model is trained to do two things at once. Solve the problem. And accurately estimate its own uncertainty.

The results were striking. RLCR reduced calibration error by up to 90 percent without losing accuracy. Both on tasks the model was trained on and on entirely new ones it had never seen.

Even more interesting: standard reinforcement learning training was actively making models worse at calibration. Not just failing to help. Actually hurting it. The MIT researchers said it directly: "The models become more capable and more overconfident at the same time."

That is the part I want you to sit with for a second.

The most advanced AI training methods in the world were making models simultaneously smarter and worse at knowing what they did not know.

This is the issue our entire industry has been ignoring while sprinting toward bigger models and faster benchmarks.

Here is where this gets personal for me.

Years ago, I wrote about ternary systems in this newsletter. The idea that human thinking is rarely a clean yes-or-no proposition. There is almost always a third option.

Yes. No. Maybe.

That third one is where most of the real cognitive work happens. Maybe is not weakness. Maybe is honesty. Maybe is the space where you slow down, gather more information, ask another question, hold uncertainty without pretending you have resolved it.

For most of computing history, we have built systems that ignore the third option. Binary logic. True or false. One or zero. Even most AI training, until this MIT work, only rewarded the binary outcomes.

And here is where it lands for me, even if I cannot prove the connection scientifically.

My brain has always lived in the third option.

A lot of dyslexic and neurodivergent thinkers will recognize this. We are not always rapid yes-or-no machines. We sit with information. We hold things in tension. We see multiple possibilities at once and we are slow to collapse them into a single answer.

For most of my life, that was framed as a deficit. Indecisive. Slow to commit. Unfocused.

But what if it was actually calibration?

What if the brains that naturally live in the "I am not sure" space have been doing exactly what MIT is now teaching AI to do?

I cannot prove this is real correlation. I am not a researcher. I am a guy with a recreation degree and a lot of time spent thinking about how my brain actually works.

But the more I learn about cognitive architecture, the more I think the brains that hold uncertainty well are not broken. They are doing something the rest of the system has been undervaluing.

In Edition 339 ("Your AI Just Forgot Everything. Again."), we walked through the five-layer self-improving AI stack from Chappy Asel's research. Memory. Knowledge. Context. Skills. Self-improvement.

In Edition 343 ("Stanford Just Measured Everything"), we looked at the AI Index and the "jagged frontier" finding. Capability is spiky. Brilliant in some areas, weak in others.

This MIT work fits underneath all of that.

Because if a model cannot accurately report its own uncertainty, every layer above it gets compromised. Memory gets contaminated with false-confidence facts. Knowledge bases get filled with overconfident assertions. Self-improvement loops get trained on flawed signals.

Calibrated uncertainty is the foundation that everything else sits on.

And it is the foundation that has been quietly missing.

The same is true for human cognitive architecture. We do not have a complete map of how human memory, reasoning, and uncertainty work together. We know there is short-term and long-term memory. We know there are different processing styles. We know dyslexic brains process visually and spatially in ways that neurotypical brains often do not.

What we do not have is a clean understanding of how uncertainty itself is processed. And whether brains that hold uncertainty differently are operating in a less effective mode, or just a different one.

The MIT paper does not answer that question. But it points at something important. Uncertainty is not a bug. It is signal. A model that knows what it does not know is more useful than a model that guesses with confidence.

That principle, applied to humans, suggests the people who naturally hold uncertainty without rushing to resolve it might be more useful in certain domains than the people who resolve everything quickly.

That is a different way to think about neurodivergent cognition. And it is one I am going to keep pulling on.

This is the part I want to highlight because it surprised me when I noticed it.

I have been building a three-layer cognitive partner architecture for a while now. The DLM Three-Layer Architecture: Socratic, Strategic, Skeptic.

The Socratic layer asks better questions.

The Strategic layer makes better plans.

The Skeptic layer holds uncertainty. It says "wait." It says "we do not know yet." It says "let me check that before we commit."

That third layer has always been important to me, but I am realizing now that I built it for a reason I could not fully articulate at the time. I built it because every reasoning system needs a third option. Every cognitive architecture needs a place where uncertainty lives without being punished.

MIT just gave me the language for what I was building toward.

The Skeptic layer is the calibration layer. It is the place where "I am not sure" is allowed to exist. It is the part of the architecture that prevents confident wrong answers from cascading into worse decisions.

Three years of newsletter work, and I am only just now seeing how this one piece fits.

That is the part of building in public that keeps me coming back. Sometimes you do not understand what you are building until someone else's research names it for you.

This also lines up with the Cognitive Balance Model from Edition 332 ("A Year Ago, I Was in a Hospital Bed").

The model has three phases: Human Initiation, AI Expansion, Human Integration. Scored on the Human Guidance Index from 3 to 15.

But underneath all three phases, there is an implicit fourth thing. Honest uncertainty.

Human Initiation is stronger when the human is honest about what they do not know yet. AI Expansion is more useful when the AI can say "I am not confident in this part." Human Integration is more reliable when both sides can flag uncertainty for review.

Calibration is not a separate phase. It is a quality that runs through all three phases.

A high HGI score requires honest uncertainty at every level. Without it, the score is just confidence in motion. With it, the score reflects real collaboration.

I think this is going to become an explicit part of how I teach the framework going forward. Not just the three phases. The calibration that runs through them.

Three things. This week.

1. Ask Your AI to Show Its Confidence

The next time you ask Claude or ChatGPT or any AI for something important, add this to the end of your prompt:

Most current models do not do this naturally. But they will if you ask. And what comes back is often more useful than the answer itself.

This is a mini version of what RLCR is trying to teach. You are forcing the calibration manually because the training has not done it yet.

2. Practice the "Maybe" in Your Own Thinking

Notice when you are pressured to give a yes-or-no answer that does not have one. Notice the weight of that pressure. Notice the moments where "I am not sure" would actually be the most accurate response.

Practice saying it out loud. To yourself. To your team. To your AI.

For a lot of dyslexic and neurodivergent thinkers, this comes naturally and gets trained out of us by environments that do not reward it. Reclaiming that third option is part of working with how your brain actually wants to operate.

3. Add a Skeptic Layer to Your Single Source of Truth

If you have built a Single Source of Truth from Edition 329, add a section called something like "What I Am Uncertain About Right Now."

List the strategic decisions you are sitting with. The questions you have not answered. The directions you have not picked. Update it regularly.

This becomes the calibration layer of your own cognitive partner setup. Your AI gets to see not just what you know, but where you are still working things out. That changes the quality of the collaboration in ways that surprised me when I started doing it.

We are at a moment in AI history where the most important findings are not always about new capabilities. Sometimes they are about acknowledging what the capabilities are missing.

MIT just acknowledged that the most advanced reasoning AI models in the world have been trained to be confidently wrong. And they fixed it with a single change to the reward function. A 90 percent reduction in calibration error without losing accuracy.

That is huge.

For our community, the lesson is bigger than the technical fix.

Uncertainty is signal, not weakness.

For the brains that have always lived in the third option, that is validation. Not just emotional validation. Architectural validation. The cognitive style that has been undervalued is the same style that AI systems are now being deliberately trained toward.

We have been doing this all along. We were not broken. We were calibrated.

And the world is starting to catch up.

Matt "Coach" Ivey Founder, LM Lab AI | Creator, The Dyslexic AI Newsletter

Dictated, not typed. Obviously.

🧠 MIT's CSAIL lab just published research showing that the most advanced AI training methods have been making models simultaneously smarter and worse at knowing what they do not know.

🎯 Their fix: RLCR (Reinforcement Learning with Calibration Rewards) trains models to estimate uncertainty alongside their answers. It reduced calibration error by up to 90% without losing accuracy.

⚖️ The deeper insight: a model that says "I am 95% sure" when it is right half the time is more dangerous than one that gets the answer wrong outright. Calibration is the foundation everything else sits on.

🔀 Ternary thinking matters. Yes, no, maybe. The third option is where most real cognitive work happens. Most computing history has ignored it.

🧩 For dyslexic and neurodivergent thinkers: many of us have always lived in the "I am not sure" space. What gets framed as indecisive may actually be calibration. The cognitive style being undervalued is the one AI is now trained toward.

🏗️ This connects to the DLM Three-Layer Architecture (Socratic, Strategic, Skeptic) and the Cognitive Balance Model. The Skeptic layer is the calibration layer. Honest uncertainty runs through all three phases of the HGI.

🛠️ Three things to do this week: ask your AI to show its confidence, practice the "maybe" in your own thinking, and add a "what I am uncertain about" section to your Single Source of Truth.