Edition 375 | August 21, 2026 The Dyslexic AI Newsletter by LM Lab AI

Reading Time: 10 minutes | Listening Time: 13 minutes

Last edition I said we need proof of human.

Then I went and looked at what's actually being built, and found something stranger than I expected.

They're building it for the bots first.

I used to talk about an "agent internet" as a thought experiment. A layer on top of the web where agents talk to agents, negotiate, buy things, exchange machine-readable information, while humans stay on the human-facing layer.

That's not a thought experiment anymore. It's a construction site.

Cloudflare has spent this year laying down infrastructure specifically for a world where AI agents are a primary workload. Compute, storage, memory, identity, permissions, browser access. There's a protocol called Web Bot Auth, currently working its way through standards bodies, that lets a website cryptographically verify which agent is knocking. Not "is this a bot," but which bot, signed and provable. Visa built it into their agentic payment platform. Mastercard put it into Agent Pay. American Express is in too. AWS, Vercel, Shopify and Akamai have implemented support.

Agents are now being issued wallets and identities so they can shop on your behalf.

And here's the number that stopped me. Cloudflare's CEO said in March that he expects bot traffic to exceed human traffic by 2027. AI bot requests across their network are already running over ten billion per week.

Ten billion a week. Next year they outnumber us.

We are building a verified identity system for software agents before we've built one for people.

I want to be fair about why. It's not sinister, it's just incentives. A merchant has a burning, expensive, immediate problem: is this shopping agent a real Visa-backed customer agent, or a scraper wearing its name? There's money attached to that question, so it got answered fast.

Nobody has a Visa-sized financial incentive to verify that the person who wrote the article you're reading actually thought about it.

So that layer doesn't exist yet.

Here's the good news. People are working the human side.

Researchers from MIT, OpenAI, Microsoft and others published a proposal for personhood credentials, a way to prove you're a real human online without revealing anything else about yourself. You do something offline that only a person can do, get issued a credential, then prove that fact cryptographically without handing over your identity. Other projects come at it differently: World ID with biometrics, Gitcoin Passport stacking signals, BrightID using social graphs.

The field has a name, proof of personhood, and real money behind it.

And there's a line in this research I haven't stopped thinking about.

Turing imagined a judge trying to tell a machine from a person. In 2026 the question flipped. The hard part isn't whether AI can pass as human. The hard part is proving that you aren't one.

Read that twice. That's the world now.

So proof of human matters, and I want it built. But I've been sitting with it for a week and here's where I landed:

Proof of human verifies that a person was present. It does not verify that a person thought.

A fully verified, credentialed, biometrically confirmed human being can sit down this afternoon, press a button, and publish ten thousand AI-generated articles. Every one of them carries a valid proof of human. Every one of them is exactly the sludge I was complaining about last edition.

Proof of personhood defeats bot armies. It does nothing about a real person operating a content factory.

That's the half it doesn't solve.

So stop asking "is this AI?" It's one question doing three jobs badly.

Was a human present? That's proof of personhood. Being built now, well funded, real progress.

Did a human think? That's Cognitive Provenance, from last edition. Nobody is building this.

Who is this for? Human-facing, agent-facing, or content that exists only to feed ranking systems and other machines. Being built, but for commerce, not for truth.

Those three answers describe completely different objects. A verified human publishing AI slop passes the first and fails the second. A dyslexic kid dictating an essay fails a detector and passes everything that matters.

Right now we've got funded answers to question one and question three, and silence on the only one that tells you whether a mind was involved.

Here's the part I want to be careful with, because it's the reason I care about any of this.

When people get fed up, and they will, faster than anyone expects once bots outnumber us, the demand won't be for nuance. It'll be for purity.

Human-made only. No AI. Certified human-written. A little badge that means nobody touched a machine.

I understand the impulse. On a bad scrolling day I feel it myself. But think about what that label would actually do.

It would lock out the people who need these tools most.

A dyslexic writer using AI to clean up dictation gets no badge. A blind user with a screen reader and an AI describing images gets no badge. A person with a tremor who dictates instead of typing, someone with a brain injury using AI to organize their thoughts, an immigrant writing in their second language with help. None of them get the badge.

Meanwhile a verified human who commissioned nothing but a topic and hit publish on ten thousand articles? That person can absolutely be certified human-written, because a human account was present the whole time.

A purity label would punish accessibility and let content farms walk right through, because it's measuring the wrong thing.

This is exactly the mistake school already made with us. The rule was never really about learning. It was about the appearance of unaided work. So the kid who needed the accommodation got treated as suspect, and the kid who cut corners cleanly got an A.

We are about six months from making that same mistake at internet scale.

Provenance is the alternative to prohibition. It's the only version of this where a dyslexic person doesn't have to choose between the tools that give them their bandwidth back and the right to be counted as an author.

That's not a side benefit of Cognitive Provenance. For us, it's the entire point.

Right now the internet ranks things by likes, shares, followers, backlinks, citations, engagement, authority.

Every one of those is now cheap to manufacture.

The signal that's actually getting scarce is human cognition. So let's name it.

Human Signal. Not human-written. Human-originated. Those are not the same thing, and the gap between them is where this whole argument lives.

A person can write every word of something themselves and contribute no original thought. Another person can hand off eighty percent of the language and supply the entire intellectual architecture. The first one passes any detector you throw at it. The second one gets flagged.

Word count was never the measure. It just happened to be the thing we could count.

Here's what I'd want, and I don't think it's far-fetched.

A toggle. In a browser, a search engine, a feed, or in your own cognitive partner. Flip it and the internet re-sorts.

Show me content where the thinking traces to a person.

Exclude anything generated entirely by agents unless a human verified it.

Find me the strongest original human perspectives on this, not the summaries of summaries.

Trace this claim back to the first identifiable human who made it.

That last one might be the most valuable thing on this list. Watch a claim move now: a person thinks something, an AI summarizes it, another AI remixes the summary, an agent turns it into a post, a site publishes it, search indexes it, and an AI hands it back to you as fact. Six transformations. The original human is technically still in there. Good luck finding them.

Property has title chains. You can trace a deed back through every owner. Information doesn't have one, and we're about to badly need it.

And notice what Human Signal Mode isn't. It's not a ban. Sometimes I want the AI summary. It's faster and it's fine. Sometimes I want a machine-generated report, and that's a legitimate thing to want.

It's not about banning synthetic content. It's about knowing which one you're holding, and getting to choose.

Same limitation I named last edition, and it hasn't gone away.

Provenance is self-reported. Anyone can claim high human origin on something a machine wrote. Cryptography solved signing a file. Nobody can sign a thought.

Which means Human Signal Mode as I've described it doesn't exist and might not, at least not soon. It needs a standard, adoption, and some way to handle the fact that the loudest claims of "human-originated" will come from whoever is doing the least thinking.

I'm not announcing a product. I'm pointing at a hole that a lot of well-funded people are currently building right past, and saying that the hole is the part that matters for people like us.

Ten billion agent requests a week says somebody's going to have to fill it.

I'd rather we were in that conversation early than get handed the finished rules and told to comply.

Two things, and the first one takes a minute.

Notice what you can't verify. Next time something makes you angry or certain online, try to trace it to a person. Not the account that posted it, the human who first thought it. Most of the time you can't get there, and noticing that is the whole exercise. That instinct is going to be a survival skill.

Then decide what you're going to say about your own work. If you use AI and you're vague about it, you're going to get lumped in with the content farms when the backlash comes. If you're specific, here's what I thought and here's what the tool did, you're building the thing I'm describing, one piece of work at a time, without waiting for a standard.

Specific beats defensive. Start now, while it's still a choice instead of a requirement.

"I'm reading something and I want to trace it back. Here it is: [paste the piece]. Help me work out what's actually original to this source versus repeated from somewhere else. What claims does it make that aren't sourced? Where would I look to find the earliest human who made this argument? And be honest with me about which parts of this you can't verify at all, rather than guessing to be helpful."

That last clause matters. A tool that guesses to please you is worse than no tool, and most of them will guess unless you tell them not to.

I did a lot of looking for this one. Go read them yourself, don't take my word for it.

On agent identity and the agentic web

The traffic numbers, bots exceeding humans by 2027 and ten billion AI bot requests a week, come from Cloudflare's own CEO about Cloudflare's own network. Treat the prediction as a prediction from a company with a stake in it.

On proof of personhood

From last edition, still relevant

Matt "Coach" Ivey

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

Dictated, not typed. Obviously.

Topic coming soon. Stay tuned.