AI Keeps Losing Your Train of Thought

Standard LLMs are neurologically incompatible with ADHD

Feb 18, 2026

🧠 Day 45 of 100. Here’s what I found while building, and why it changes the architecture of everything.

Let me tell you what’s actually happening when an AI gives you a response.

It’s not retrieving facts. It’s not following your instructions step by step.

At every single token, it’s running a probability calculation. ‘What word is most likely to come next?’ And then it picks that word. And then it does it again. And again.

There’s a name for what’s happening here. Let’s call it Probabilistic Drift.

I’m Hodman Murad, Founder of Asaura AI.

For the past several weeks, I’ve been documenting, in public, the biology behind why ADHD brains stall out at work.

I wrote about stall, the moment your brain hits a cognitive load threshold and stops cooperating entirely. I wrote about the planning tax, the mental energy burned just figuring out how to start, before a single word gets written. I wrote about working memory’s hard biological limit and what happens when a vague task requires your brain to hold 15 decisions at once. And I wrote about the recency and sequencing engines I’m building into Asaura to offload that executive function work externally.

All of that research pointed to one underlying problem. As currently designed, AI tools worsen all of these issues.

Why ‘Most Likely’ Is a Trap

Here’s the problem with building tools for ADHD brains on top of probabilistic systems.

Your brain is already fighting a drift problem.

Executive dysfunction means the signal degrades mid-task. You start on one thing, the thread frays, and suddenly you’re three tabs deep into something completely unrelated. You didn’t choose that. Your working memory just... let go.

Now imagine your AI tool does the same thing.

You give it a clear instruction. It starts well. Then, because ‘most likely’ tends to pull toward its training data rather than your specific intent, the output drifts. It adds caveats you didn’t ask for. It restructures your request. It answers a slightly different question than the one you wrote.

For a neurotypical user, that’s mildly annoying.

For an ADHD user, it breaks the entire cognitive loop. You have to re-read, re-orient, re-anchor. And by then, the moment is gone.

🔬 Building on What the Research Shows

Building Asaura AI has meant getting close to how constrained decoding actually works. One concept in particular reshaped how I think about the architecture as a whole.

Here’s what’s actually happening:

Standard AI generation is a free-for-all at every step. At every token, the model scores its entire vocabulary and picks the most probable option. Nothing stops it from drifting toward a generic, slightly off target answer.

Constrained Decoding changes that at the infrastructure level. Instead of a free choice, you build a Finite State Machine (FSM) around the generation process. At every step, invalid tokens get masked out entirely. The model can only move forward through states that have been pre-defined as valid.

Architectural constraint removes the condition that makes drift possible in the first place.

Conversational AI Was Never Built for This

Most AI tools are built for conversation. You ask, it answers. The format is loose. The output is flexible. That’s fine for many use cases.

But ‘conversational’ is exactly the wrong paradigm for an ADHD productivity tool.

Conversation is open-ended. ADHD brains need structure.

FSM-based decoding gives you something conversational AI was never designed to offer: a generation process that’s accountable to a structure before it produces a single word.

Instead of ‘reply with something helpful,’ the system operates like this:

Step 1. Here is the schema of a valid response.

Step 2. At every token, only tokens that keep you inside that schema are permitted.

Step 3. Generate until the schema is complete. No more, no less.

A system built this way isn’t trying to be helpful in a general sense. It operates within a defined boundary, with every output verified against a structure before it reaches you.

What This Means for You

ADHD users deserve tools that don’t make the executive function problem worse.

You shouldn’t have to re-read AI output three times to check if it actually answered your question. You shouldn’t have to correct the model when it goes off-track. You shouldn’t have to do cognitive cleanup on your own productivity tool.

Structural AI, built on FSM-based decoding, means the system is accountable to a format before it generates a single word.

That’s where we’re headed.

Later this week, we’re going to get into why structure is faster, not slower, and what that means for building tools the ADHD brain can actually trust.

💬 Reply to this with one word: what does AI-drift cost you most often? Focus. Time. Trust. I want to know.