Edition 374 | August 21, 2026 The Dyslexic AI Newsletter by LM Lab AI

Reading Time: 10 minutes | Listening Time: 13 minutes

Friday morning. Coffee's on, most of the week is behind me, and I sat here scrolling for a solid ten minutes before I ever picked up the microphone.

That scroll is what this whole edition turned out to be about.

I want to start by admitting something that doesn't help my brand.

I'm getting tired of it.

I scroll now and I can't tell. Is a person behind this post, or a bot? Did anyone actually watch the thing this video is reviewing? Was this article written by a reporter or assembled by an agent from three other articles that were also assembled by agents? Even the news is getting slippery, and that one bothers me most.

And I need you to sit with who's saying that.

I'm the AI guy. I've written 374 editions defending this technology. I use it every single day. I've spent two years and most of my savings building on top of it. If anyone was going to have the patience for a feed full of synthetic content, it was going to be me.

If I'm frustrated, then the people who never wanted any of this are somewhere well past frustrated.

That's the part I keep coming back to. I've been talking to skeptics for two years about the upside. Meanwhile, the actual daily experience of the internet has been getting worse in a way that proves them a little bit right, and I owe it to them to say so out loud.

But I don't think the answer is the one everybody reaches for.

The reflex is to ask: "was this written by AI?"

That question sounds smart and it's nearly useless. Let me show you why.

Picture four people writing an article on the same subject.

Person A asks AI to write an essay about something they know nothing about. Reads it once, changes a word, publishes.

Person B spends two hours working out an original argument, dumps the mess into AI, and asks it to organize the thing into something readable.

Person C records a twenty-minute voice memo about an idea they've been chewing on for months, then uses AI to turn that transcript into an article.

Person D writes every word themselves, but leans on spellcheck and grammar correction the whole way.

Run all four through a detector and you get noise. B and C might score more "AI" than A, because A edited theirs and they didn't. D might get flagged because grammar correction sands off the human tells.

The detector is looking at the words, and the words are the last and least interesting thing that happened.

Person A supplied nothing but a topic. Person C supplied everything except the sentence construction. Those are opposite acts. The finished pages look like cousins.

So the question isn't whether AI touched it. The question is: where did the thinking come from?

I'm calling that Cognitive Provenance. And I have to be honest with you about where it came from, because it isn't new.

Back in Edition 332 I introduced the Cognitive Balance Model to you. Human Initiation, AI Expansion, Human Integration. Three phases, and a scoring system underneath it called the Human Guidance Index.

That work had a piece I'd honestly forgotten about. Buried in the supporting metrics of the white paper is something I named the AI-Originality-Human Score, and its entire job was to separate two things: translation, where AI amplifies a human insight, and generation, where AI produces the insight with barely a human involved.

Translation versus generation.

That's this whole edition. I wrote it down two years ago in a formula, put it in a white paper, and then talked myself into thinking I'd had a new idea last week.

I only caught it because I went looking through my own research this morning.

I don't think that's embarrassing. I think it's the strongest evidence I have that this distinction is real. I arrived at the same place twice, from two completely different directions, two years apart, without remembering the first trip. Ideas that hold up tend to do that. The ones that don't, you have to keep propping up.

So this isn't a new framework. It's the same argument I've been circling since 332, finally pointed at something specific.

Provenance is an old word from the art world. It's the documented chain of custody for a painting. Who made it, who owned it, where it hung, what got restored and when.

Provenance never asked "is this a real painting." Obviously it's a real painting. It asked where it came from and what happened to it along the way.

Stop authenticating the artifact. Start tracing the process.

Applied to anything we make now, it separates things a detector mashes into one blurry score:

Human-originated thought. The ideas, observations, experiences, hunches, connections that started with the person.

Human-directed exploration. The questions they asked the machine to push, challenge, or extend their own thinking.

AI contribution. Actual new substance the machine introduced. Counterarguments, examples, framing they didn't have.

AI transformation. Structure, grammar, spelling, punctuation, organization. The mechanics.

Human judgment. What they kept, killed, corrected, or refused.

Look at Person A and Person C through that and they stop resembling each other immediately. A has almost no human origin and total AI contribution. C has total human origin and almost pure AI transformation.

Same output category. Opposite provenance.

Inside that, there's a distinction I think matters enormously, and almost nobody makes it.

AI substitution. I didn't have the idea. The machine generated the idea and wrote it.

AI augmentation. I had the idea. The machine helped me develop and express it.

AI accessibility. I had the idea and could explain it out loud fine. The machine removed the mechanical barrier between my thinking and readable text.

Those three can produce nearly identical documents. They are not remotely the same act.

Right now we call all three of them "AI-generated content," which is like calling a ghostwriter, an editor, and a pair of glasses the same thing because a book came out of all three.

Here's what actually happens when you make me type in real time.

I have a thought. I hit a word I can't spell. I stop. I go find the spelling. And when I come back, the thought is gone. Not weaker. Gone. I've lost the thread I was following, and now I'm spending energy trying to recover a thought instead of having the next one.

That's not a spelling problem. That's a working memory tax, charged on every sentence.

The old loop looked like: think, type, misspell, stop, correct, reread, try to remember what I was saying, continue.

The loop now is: think, speak, keep going.

Talking removes the tollbooth. I can talk through an idea far faster than I can type it, and I never have to stop. The raw transcript is a disaster. Wrong words, no punctuation, sentences crashing into each other. But the thought is intact, and the machine can reconstruct what I meant from the wreckage.

The AI is not making me smarter. It's giving me back the bandwidth I was spending on spelling.

Every word of this newsletter came out of my mouth. It always does. That's what the sign-off has been telling you for 374 editions.

Call that "AI-generated content" and you've described the least important thing that happened.

Here's what surprised me when I went looking.

We already built a content provenance standard. It's called C2PA, and the consumer version is Content Credentials. Adobe, Microsoft, the BBC, Intel and Sony started it back in 2021, and it now has thousands of member organizations behind it. High-end cameras from Leica, Sony, Nikon and Canon cryptographically sign photographs at the moment of capture. OpenAI and Google aligned on it this past May. Newsrooms sign their photos with it.

And as of this month, the EU's transparency rules kick in, requiring machine-readable disclosure on AI-generated content.

So this isn't fringe. The world already quietly agreed that provenance beats detection.

But look at what that infrastructure actually tracks.

It tracks the file. Which device made it. Which software touched it. Whether the bits changed after signing.

None of it tracks the thinking.

A credential can tell you an image came out of a specific camera and went through Photoshop twice. It can't tell you whether the photographer understood what they were looking at. The standard's own stated limitation is that it proves a claim was made, not that the claim is true.

We built a beautiful chain of custody for pixels and skipped the layer where the human actually lives.

And that gap gets worse fast. Watch what happens to a single idea now: a human thinks something, an AI summarizes it, another AI remixes that summary, an agent spins it into a post, a site publishes it, search indexes it, and an AI hands it back to you as an answer.

Six transformations. The original human is still technically in there somewhere.

Try tracing it.

Without provenance, that lineage is just gone. With it, you'd have something like a title chain for information, the thing that lets you follow a claim back to the first person who actually thought it.

Here's where this stops being a framework and becomes a kid.

Right now a dyslexic student turns in an essay and the teacher sees exactly one thing: the final document. Then the teacher has to make a judgment call with almost no information. Either the writing is clean and they wonder if a machine did it, or it's rough and the grade quietly measures mechanics instead of thinking.

Either way, that kid's ability to think is being judged through the one channel their brain handles worst.

Now imagine the student hands over the provenance instead of just the artifact.

The voice recording. The messy raw transcript. The questions they asked. What the AI suggested. What they rejected, and why. The research they went and found themselves. The revision trail. The final version.

The teacher stops grading a document and starts watching a mind work.

And the student's spelling stops getting mistaken for the student's thinking.

I don't know how to tell you how much that would have changed things for me. Or for my kid at that table.

That's not a detector. That's not surveillance. That's the process finally being visible to someone who was previously handed an output and asked to guess.

I'm going to break my own idea before somebody else does me the favor.

Provenance like this is self-reported. I could claim high human origin on something a machine wrote start to finish. Nothing stops me. The moment any of this carries weight, a grade, a job, a search ranking, people will game it, and the loudest claims of "human-originated" will come from whoever is doing the least thinking.

C2PA got around that with cryptography, because signing a file at the moment of capture is a solvable engineering problem. Signing a thought is not. There's no camera pointed at the inside of your head.

So I don't think the honest version of this is a score.

Which is awkward, because I already publish four of them. HGI, CLR, CPAS, AOH. All with formulas. All in the white paper. My first instinct here was to add a fifth, and I had it built: human origin 95, cognitive partnership 91, transformation high. It looked great.

Then I worked out why my existing scores are fine and that one wasn't, and the difference matters more than the scores do.

My white paper metrics are a bathroom scale. A published provenance score is a drug test.

HGI is something you run on yourself to get better at working with AI. Nobody's trying to beat it. There's nothing to win. If it's off by a point, you've lost nothing.

A provenance score attached to a published artifact is adversarial the day it means anything. Put a grade on it, a job, a search ranking, and people optimize against it immediately. A number that looks defensible and isn't becomes a weapon, and it gets pointed at exactly the people who need the tools most.

Same math. Completely different situation. I should have made that distinction in the white paper and I didn't, so I'm making it now.

What I think this actually is, at least today, is an artifact you choose to show. Not a number assigned to you. A trail you keep and hand over when it matters, the way a painting comes with its papers.

Which means it works in the places where somebody is willing to look at the trail. A classroom. A hiring conversation. A client relationship. It does not scale to the entire internet, and I'm not going to pretend otherwise.

Smaller claim. But it survives contact.

Look at what these have in common.

Edition 332 said stop judging a collaboration by the output, look at how the human framed and refined it.

Edition 370 said stop measuring yourself on a scoreboard that was never counting your actual work.

Edition 373 said stop measuring whether a student studied, look at evidence their capability changed.

This one says stop asking whether AI touched the output, look at where the thinking came from.

That's the same sentence four times.

We keep grading artifacts because artifacts are easy to grade, while the thing we actually care about, did a mind do something real here, leaves its evidence in the process.

The trail is the proof. It was the proof in the white paper, it was the proof in school, and it's the proof on the internet.

I didn't plan that. I just kept talking into a microphone for two years and it turns out I've been saying one thing the whole time.

You don't need a standard to exist. You need the habit.

For one thing you make this week, keep the trail. The voice memo. The ugly first transcript. The prompt you actually used. The version you threw away. That's two minutes of saving, not a system.

Then say something true at the bottom of it. Mine has been one line for 374 editions and I never once called it a provenance statement until this morning.

And if you're a parent or a teacher: ask to see the process once. Not to catch anybody. Ask a kid to walk you through how they got there, what they tried, what the tool gave them, what they refused. You'll learn more in five minutes than the finished page was ever going to tell you.

"We just finished working on this together. Help me write an honest provenance note for it. Walk back through our conversation and separate out: which ideas and arguments originated with me, what I asked you to explore or push back on, what substance you introduced that I didn't have, what you only changed about structure and language, and what I rejected or overruled. Be accurate, not flattering. If I contributed less than I think I did, say that plainly."

That last sentence is the whole thing. If it comes back and tells you the machine did the thinking, that isn't a failure. That's the tool working.

New thing starting today. When an edition uses outside research, I'm going to show you what I used so you can go read it yourself instead of taking my word for it.

Matt "Coach" Ivey

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

Dictated, not typed. Obviously.

Everyone's racing to build "proof of human," credentials that verify a real person is behind an account. It's real, it's funded, and it solves the wrong half of the problem. A verified human can publish ten thousand AI articles before lunch. Next edition: why proving a person exists isn't the same as proving a person thought.