The conversation about AI has gotten very loud. So far, in that conversation aspects about the human experience in the AI era have gotten a bit lost.

The Event Horizon is a Thursday letter built for founders who want to look clearly at what AI is actually doing—to our work, our thinking, our identity, and our businesses—and lead from that clarity rather than react from the noise.

We’re not here to chase the tools or debate the hype. We’re here to figure out what it means to stay human while building something powerful—and to do that work together, week by week.

Human lead. AI empowered. Always.

Welcome back to The Event Horizon. 🌅

Two issues ago, I told you to hang onto a line: the training data is us.

This week, that line runs out. AI is already running out of human-written material to learn from—and the fix the industry has already started reaching for is AI teaching itself off its own homework.

AI teaching itself off its own homework is a bit like a rumor citing itself as a source—or spellcheck learning spelling from your typos folder. Either way, the system isn’t being corrected by reality. It’s being trained by its own remnant.

There is a huge debate raging around the how, who, why, and what LLM’s should or shouldn’t train on. It is, without hyperbole, an extremely nuanced and complex conversation, to be sure.

What I’d like to talk about today is the actual path in front of us right now, possible forks in the road, and a question at the end—instead of an offered solution.

Because something that involves the majority of humanity is not something we solve in an essay.

That only happens with continuing the conversation, and adding ideas to the collective brain until each expression of the argument is not only addressed but also reasoned out and considered.

This letter is merely one open path to continue that conversation.

📡 The Signal

Twenty Generations From the Original

What do twenty generations, seven photocopy machines, four distinct signatures, three density settings, and one forensic document examiner have in common?

No, this isn’t a “walked into a bar” joke, and it’s not some bizarre “12 Days of a techno Christmas” song, either.

In 2017, a forensic document examiner by the name of Robert Gervais ran a rigorous study on photocopier distortion. He took four genuine signatures—he wrote two each with two different pens, and at two different speeds with each pen, giving him the original four samples to work from. Then he ran those originals through a collection of photocopying machinery of varying capabilities and at different setting capacities. Some were analog, some were digital, and different makes and models, testing different levels of technology—seven in all.

For each original handwriting sample, he then reproduced twenty generations—in other words, using the previous copy to make the next copy generation. In total, he produced 1680 images for this study.

Here’s what he found: the first four generations of any of the series of copies didn’t register obvious distortions to the naked eye, even through the software the series was uploaded into to examine the images. They looked clean, and mostly indistinguishable from the original, even on machines that go on to produce severe distortion.

By generation twenty, however, some are veritably unrecognizable.

The damage doesn’t announce itself early. It compounds quietly, underneath a surface that still looks trustworthy—and by the time it’s visibly wrong, you’re many generations past being able to point to where it started.

Here’s why I’m telling you about a paper photocopier study in a newsletter about artificial intelligence.

Large language models (LLMs) are continually trained on data. That part isn’t news to anyone reading this. What’s newer is what that data is increasingly made of.

For years, the raw material was us—the accumulated, uneven, occasionally brilliant mess of everything humans have written, photographed, recorded, and put online. The original signature, so to speak. But AI-generated text, images, and audio have been flooding back into the same pool they were scraped from in the first place. Which means the next generation of models isn’t only learning from the original anymore. It’s learning from copies of copies—its own prior output, filtered back in as though it were fresh material.

Same mechanism. Different medium.

A model trained partly on a previous model’s output is generation two. Trained on a model trained on that output, generation three. And so on.

The photocopier never cared whether it was reproducing a signature or a sentence. Neither does the AI model.

And the mechanism itself isn’t new, either. It existed before AI came into full view. AI didn’t invent generational distortion. It just found a way to run it a lot faster, on a lot more material, with a lot less time between generations to notice.

The Well Is Already Running Dry

So, if the technology is still so new, why does AI even need to train on its own data?

Let’s start with the research from Epoch AI that estimates available new high-quality public human text runs out somewhere 2026 to 2032. You read that right: the low end of the window is now.

How did they arrive at that timeframe?

As the authors begin in their report, “Scaling has been a key factor driving progress in AI. Models are growing in parameters and being trained on increasingly enormous datasets, leading to exponential growth in training compute, and dramatic increases in performance. For example, five years and four orders of magnitude of compute separate the barely coherent GPT-2 from the powerful GPT-4.”

Basically, every time the model evolves into its next iteration of capability, its appetite for new datasets increases.

Researchers who track this for a living estimate a point inside the above window as when the usable stock of human-written material online gets fully used up in past training datasets, at the pace things are currently going.

But the pace isn’t a fixed variable, either.

Some companies are already feeding models data faster than it naturally accumulates, just to squeeze out more performance sooner.

So, here’s the number underneath the window: pushing that hard, the target inside the window collapses down to 2027. Push harder still, with more aggressive scaling strategies, and it collapses again.

And according to the Communications of the ACM, that target date isn’t actually somewhere off in the even near future… It’s already here.

The CACM piece earlier this year, “Model Collapse Is Already Happening, We Just Pretend It Isn’t,“ confirmed collapse is measurable in production systems now. Through background-removal software failing on specific hair textures, image generators drifting toward homogeneous output, they show that this is not a future risk anyone can still treat as theoretical.

Sound familiar? That background-removal glitch, the subtle drift toward sameness in image generators—that’s generation four. Recognizable. Still trustworthy-looking. Which was exactly the problem with the signature, too.

By the way, Alex Williams, the author of the CACM article, likes my photocopy study for this, too (though he doesn’t know it). He uses the metaphor, as well, for what model collapse looks like.

Model collapse isn’t a future risk anymore, and the fix everyone’s reaching for is what’s causing it.

Downstream From the Well

Generation twenty isn’t the only problem. Nobody’s even agreed yet on what generation one was even allowed to be copied from.

While one clock is running forward, toward a shortage nobody’s solved yet, a second clock has been running backward this whole time. Nobody’s actually settled how the current generation of models got built in the first place.

2026 denotes three years since the separate reckoning of the second clock started over the data already used to build the current generation of models through (still) active lawsuits. The unsettled legal fight between The New York Times v. OpenAI and Microsoft over copyright infringement doesn’t look any closer to resolution.

And Anthropic is in the ring, too, with an actual “after the fact” lesson on what the price of consent really is. At the start of building LLMs, companies were using books to train the language models, before moving on to the whole of the internet.

In June 2025, the judge for the case a class of authors brought against Anthropic for copyright infringement actually split the verdict. He ruled that the training itself was legal, but where the books came from and what Anthropic kept was not. The court ruled a settlement of $1.5B be paid by Anthropic for almost 500,000 works it used to train Claude on.

The industry is still settling the past whilst mortgaging the future in the same news cycle. The industry calls it a data shortage. What’s disappearing is harder to license back than a database.

If the well is running dry, and what’s refilling it is increasingly the machine’s own reflection, what happens to the next model that trains on it, and to the humans reading, writing, and thinking inside the same polluted supply?

🪨 First Principles

Every Machine Distorts Differently

We have now come to the fork in the road… or rather what I currently see as four forks in the road. Four possible paths through this training data challenge in front of us.

Paying For What Was Already Taken: Fork one, retroactive licensing

In this scenario, we collectively choose to settle for what was taken, and the models keep training on scraped material, only under new terms.

Like the authors in the Bartz v. Anthropic case above, compensation would mostly come retroactively on the material these LLMs have already been trained on, with the knowledge that the compendium of fresh human-written text is dwindling to a pace that cannot keep up with the pace of the models expansion and training need.

Consent becomes an “after the fact” settlement, not a starting condition.

Ask Before Taking: Fork two, consent-first sourcing

Last fall, Warner Music Group (WMG) partnered with Suno, an AI music generator, striking what they’ve called, “a first-of-its-kind partnership.”

In the Warner–Suno’s model, AI systems are built only on material artists knowingly gave consent to be used, with ongoing control retained.

WMG’s CEO Robert Kyncl said of the development, “AI becomes pro-artist when it adheres to our principles: committing to licensed models, reflecting the value of music on and off platform, and providing artists and songwriters with an opt-in for the use of their name, image, likeness, voice and compositions in new AI songs.”

So, as this model demonstrates, this scenario is slower, and a bit smaller, but real consent—instead of retrofitted consent.

Betting The Mirror Won’t Crack: Fork three, synthetic data at scale

Following the path fork number three lays out, we collectively accept the shortage. We also accept the degradation risk that we talked about in The Signal, which, as you remember, is already occurring. The idea would be to optimize around it as much as possible.

At the moment, “optimizing datasets” is the main alternative lever to raw scaling. Instead of asking “how much bigger, how much more,“ it starts asking “what exactly are we feeding it, in what order, and from where.“

What would that look like in real life?

Well, Microsoft has already “optimized” this process, deliberately, and it worked better than they anticipated. They built a model called Phi-4 almost entirely on data it generated itself rather than scraped from the internet. Roughly 50 different types of made-up training material, built specifically to teach the model different skills.

And it outperformed larger models that were trained the traditional way, on real, human-written material.

This is the clearest real-world proof that Fork 3 isn’t just a risk, it’s a strategy some companies have already bet on and won with, at least so far.

Anthropic’s actually been doing a version of this since 2022.

Their models help train themselves by critiquing and revising their own answers, and those self-critiques become part of what teaches the next version what “good” looks like. It’s older than the current shortage conversation, but it’s a clean example of a company choosing, on purpose, to let a system learn from itself rather than treating that as an accident to avoid.

This isn’t hypothetical, as it is increasingly the answer companies who build and train LLMs are reaching for. So while it is a viable answer to the well drying up, it is also the biggest culprit to the echo beginning to happen in the output of the models.

A hall of mirrors, producing quality drift nobody chose, absorbed without fuss, fight, or force, generation over generation.

The Door Nobody’s Built Yet: Fork four, collective infrastructure not yet built

While there isn’t a lot of building in this area yet, there is quite a bit of research, policy work, and even one viable example. So let’s take a look…

The whole idea already has a name, and someone gave it one three years before this became an emergency.

In 2023, researchers formally proposed a “public data trust,” a mechanism to hold and license AI training data on behalf of the public, back when the shortage was still a forecast, not a headline.

They wrote up the blueprint for doing this at real scale. The academic paper proposed an organization that would hold and manage training data on behalf of the people who created it, so individuals aren’t negotiating one-on-one against companies with far more power and lawyers than they have.

A separate peer-reviewed study asked the harder question: would this actually work?

Author Pamela Samuelson didn’t ask “is this a nice idea?“ Instead, she look head-on at the more real-world challenge of “is collective licensing genuinely feasible at the scale AI training actually happens?“

So, serious researchers are testing the idea against reality, not just proposing it as a lovely premise.

Even with the proposed challenges, a government is already trying to build this, not simply discuss it.

In March of this year, the European Parliament adopted an actual system for this idea. A legal framework to clarify licensing rules for generative AI and support voluntary collective licensing agreements between companies training AI on copyrighted creative work and creative industries. The EUIPO, the EU’s intellectual property office, would facilitate, as a trusted intermediary, management of opt-out registries, and help broker licensing deals. It also explicitly calls on the European Commission to look at fair compensation for past uses, not just future ones.

Real institutional movement towards the fourth fork’s path, not just theory.

And as it happens, someone already built a small, working version of this.

Fifteen choirs in the UK got together and set up their own system for deciding, collectively, whether—and how—their recorded singing could be used to train an AI voice model. This is not a company deciding for them, or a lawsuit after-the-fact in play. This is the choirs themselves, governing the terms up front, together.

It’s small, and it’s only one narrow use case, but it’s real and it’s already been built.

So, data cooperatives, revenue-sharing models, and policies that could move it all forward, are being theorized on standards that don’t exist—yet. So far, it costs nothing but time and imagination, because it just isn’t real at any scale yet.

Ultimately, there’s no way to definitely say how this whole thing sorts out.

One thing is clear, however: this isn’t a future choice.

It’s a current big, hairy mess of overlapping ones, happening simultaneously across different companies, lawsuits, organizations, and platforms.

What’s genuinely useful to a conscious founder isn’t a tidy wrap-up of which path wins. As I said, there is no way to really know that yet. What’s important for you is seeing plainly that several paths are being taken at once, by different actors, with no collective agreement, and understanding what each one has costs and to whom.

🫀 The Human Layer

Are You Still Generation One?

In issue #003 of The Event Horizon, we talked about AI detectors and how turning over judgement to them begins the slippery slope to the flattened average of all human writing.

As whole, my argument across Rooted & Rising on originality is that it is a personal and brand differentiator. And it certainly is.

There’s another dimension, however, to the importance of originality that has surfaced through this topic of AI data and output distortion. As a contribution to a shared, finite resource—AI training data—what “average” will mean once the majority of output is built from AI’s own yield rather than human writing will expand exponentially. You’ll see it in the voice homogenization even for the output of human writers, too, as they increasingly lean on AI systems in their production processes.

So it won’t just be the models who are bland, lifeless, and without nuance.

If we are not careful, conscientious, and agile, it will happen to our voices, too, as a never-ending ouroboros of recycled thoughts, interpretations, and evaluations become the well everyone fills their metaphorical chalices from. As we choose to continue to write alongside the same handful of tools—without replenishing the foundation we all pull from, and without being more deliberate in how much of what we produce is our own original thought and examination—it will become harder to locate distinctiveness as everyone’s influences begin to overlap, dissolving into the gray blur of homogenization.

All while that very distinctiveness becomes more and more valuable.

This naturally leads to a discomfort of realizing your own authored thinking is now, in a very literal sense, scarce raw material, not just a creative expression.

Sitting with a question that doesn’t resolve itself is the work right now—at least it is for me.

And I think rather than reaching for false closure, the discomfort—the struggle of this big idea and complex problem—is the important thing in the long run. It raises more questions than answers, and I think it should.

Questions like:

Now that the average itself is starting to fill with AI’s own blended output, while human writing simultaneously drifts toward assistant-influenced patterns, does that make original work truly more valuable, or more exposed? Or both?

Will I notice my own drift by generation four, or only once it has already spanned out to generation twenty?

How do we ensure that these systems both stay in service to humanity and stay equitable?

Which fork in the road brings humanity to the best possible future, or is it some collaborative paths of multiple forked paths?

🌅 Light on the Horizon

Where the Signature Still Holds

A Catalogue of What’s Still Clean

Someone is keeping score of what’s still generation one, in public, for free.

Low-background Steel is a public catalogue, maintained by hand, of data sources whose contents predate or deliberately exclude the flood of machine-generated text, images, and audio that started in late 2022. Not a company, not a lab. A list: a Common Crawl snapshot from August 2022, a Wikipedia dump from the same month, Project Gutenberg, a database of word frequencies that stopped updating on purpose once its sources got contaminated, an archive of ship videos, a decommissioned coal mine in Svalbard holding a 2020 snapshot of every public GitHub repository.

The name borrows from actual low-background steel, metal forged before 1945 nuclear testing, uncontaminated by fallout, still used today in instruments sensitive enough to need it. Someone looked at the exact problem this issue has been circling and reached for the same idea: catalogue what’s provably clean, before it’s impossible to tell anymore.

It’s the plainest possible answer to “how would you even know which generation something is.” You’d have to have kept track from before the copying started. Someone already did, and it’s updated as recently as today. Read more →

Building the Thing That Tells You Which Generation This Is

If the real fix for model collapse is “taking data provenance seriously as an engineering discipline,“ someone’s already building that infrastructure, and it’s not staying theoretical.

Content Credentials, built on a standard called C2PA, attach a verifiable record to a piece of media: what made it, what it’s made from, whether it’s been touched since. Google’s Pixel 10 and Sony’s newest professional video camera both build this in at the moment of capture. OpenAI embeds it in DALL-E 3’s output. The EU’s AI Act and a California law now require exactly this kind of disclosure, and the standard is what makes it enforceable rather than a policy with no teeth.

Not perfect, and certainly not universal yet.

But it’s a real, working answer to the question this R&R edition has talked about at length: not “did AI make this,” but “can you actually trace where this came from.” It’s not a proposal. It’s already in a phone millions of people own. Read more →

⚠️ Concern Worth Naming:

Model collapse, at founder scale.

Something worth watching:
The CACM finding above names an industry-level risk.
The founder-scale version is every deferred-origination decision—letting AI draft something end-to-end rather than originating and refining as you build with it—is a small contribution to the same averaging mechanism, at the scale of one person’s output instead of one model’s training run.
Run that across enough founders leaning on the same handful of tools, and the risk stops being personal.

The good version of this looks like:
Founders who keep originating aren’t just protecting their own differentiation. They’re keeping the well from narrowing for everyone drawing from it next, human and machine alike.
That origination is the loop. Stay in it.

🎯 Your Move

You just read four forks. None of them got my verdict, on purpose.

This isn’t a “which one’s right?“ kind of essay. It’s the “here’s where we actually stand“ kind.

But I’m not going to ask you to name your fork without naming mine.

So here’s mine: Fork number four is the one I personally lean into most as what I believe will be the best path forward, both for humans and for the models. And with a working framework, a specimen to examine, and institutional movement in the arena, the future on the path that follows fork number four looks bright indeed. 😎

I could very well be wrong about that in a year. Less, even. Genuinely.

There could be forks in the road that haven’t shown themselves yet. This is moving fast enough that I’m holding it all quite loosely, on purpose, while still keeping the questions top of mind.

SO, your move this week isn’t a task.

It’s this: which fork do you find yourself standing closest to right now, and why.

Not the one that sounds best. The one that’s actually true of how you’re operating.

Then tell me, down in Signal Back. That’s not a formality this week. That’s the actual point of writing this letter, that ends with questions, instead of an essay that ends with a tidy little bow on it.

🧭 Constellation Compass

🌌 Elsewhere in the SPS ecosystem recently:

🗺️ In this issue of Elegant Email Ecosystems, This week I made the case for framework thinking: why expertise without a container survives the moment it’s delivered and disappears the moment someone tries to pass it on, and how a signature framework turns scattered depth into something a client can actually carry forward. [Read E3 #030 →]

🌱 In this issue of The Master Work, I sat with the two blueprints every business runs on, the one on paper, and the one made of everything you haven’t examined yet, and what it actually takes to catch a fear or an old story before it makes a decision you’ll mistake for strategy. [Read TMW #003 →]

🔮 Inbox Alchemy is the lab side of SPS—where ideas like these get turned into practice through a sequential, evergreen email experience—the hands-on companion to everything SPS publishes, one issue at a time. If you’re ready to move from insight to implementation, [Join us in the lab →]

📡 Signal Back

This week’s question:

Where do you land on the forks above?

I’m not looking for a polished take, just where you actually are right now.

I read every reply, and this issue especially will help me see the whole picture if you signal back.

If this issue resonated with you, I have four small asks:

① Hit the ❤️—It takes one second and tells Substack this conversation is worth having.

② Hit the 🔄 restack—It puts this in front of your followers—the ones who are already building differently and don’t yet know there’s a name for what they’re doing.

③ Share this issue with a founder you know who could benefit, and might even think you’re their hero. 🦸

④ Drop a comment—I read every one. And I reply.

Some of my best thinking happens in response to what you bring to the conversation here—and future issues often start in a comment thread (this issue is proof of that).

💡The right idea finds the right person at the right time. You might be the one who gets it there.

🌿 The Understory

What you just read is the argument. What follows is the work.

What comes next was built for The Understory, and it goes where this issue has been pointing all along: into the actual practice of staying part of what’s worth drawing from, instead of only ever drawing from it.

This week’s Understory is three things. A Practice Protocol: “The Contribution Ledger,” a simple, ongoing habit of tracking what you originated this week versus what you recirculated, continuing the ledger this letter’s been building since #003’s Rejection Ledger and #004’s Narrate the Call. A closer look at Fork Four, and why, out of four honestly-argued paths, it’s the one I’m personally betting on, the fuller case I didn’t make in the letter above, on purpose. And a Research Deep Dive: the full annotated source set behind this issue’s argument, built for the educators, coaches, and expert-to-educators in the Grove who’ll want the receipts.

And at the close—a brief, honest note on how I used AI this week. This is a standing section in every The Event Horizon Understory, because my position on AI is simple and steady, and now refined: use the tools, stay in the loop, notice the pattern—stay in the lead. These are not in conflict. The Understory is built on exactly that premise.

Let’s get to work.

For everyone reading from The Canopy: this is what The Understory looks like. Join us in the deeper work.

Join The Understory →

✨ Here’s to staying part of what’s worth drawing from, not just what draws from us.

Stay curious. Stay human.

~StacyLynn
Founder, Sitting Pretty Strategies
Root Deeply. 🌳 Rise Differently.

P.S.

If you know a writer, researcher, or founder who’s been quietly producing real, original work without realizing how much that’s actually worth right now, not just to their business, but to the well everyone’s drawing from, forward this their way. 💌

Feeding the Machine Its Own Tail