Detection Is Not Discernment
The argument about AI detection is really an argument about who holds the judgment call, and the verdict on value still belongs to you, every single time.
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 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—where we look at what AI is becoming, what it’s asking of us, and what becomes possible when humans stay in the lead! 🌅
Substack integrated an AI-detection feature last week—without any advance notice or a prior warning period—and the internet split into its usual camps: necessary trust-check, or invasive overreach.
I want to stop here to say that I genuinely had no intention of joining this particular conversation. For one thing, I think if you follow or read me at all, you know where I stand on AI at the moment. And I do not intend on using the detection tool at all myself.
My mind works just fine, thank you.
Personally, however, I don’t think either camp is asking the more important question underneath this emotionally-charged roil. So I decided this week’s TEH will. This is the third turn in a thread I’ve been building since the inception of The Event Horizon, and it lands on what I believe is the real conversation we should be having instead.
Have we handed the machine the one job that was never its to do: deciding what’s good, valuable, and worth our attention?
📡 The Signal
As was thoroughly debated and commented on all over the platform this past week, Substack integrated Pangram’s AI detector (Pangram is a detection software company founded in 2023) as a new platform feature, July 21–23, alongside releasing Substack CEO Chris Best’s essay “Against Claudefishing.” on the platform’s publication, The Substack Post.
This is worth your time and attention because the platform named its values out loud, then immediately had to live with the cost of enforcing them through someone else’s tool. Even tech news, creators across LinkedIn and other social channels, and outlets like The Atlantic got in on the discussion.
Chris Best is right that the core problem is a mismatch between reader expectation and reality—when someone invests attention in text with “no human thought on the other end.” But Substack’s answer is a detector that estimates how many words a model likely wrote.
That’s a category error.
All writing is already intertextual and cumulative: built on prior ideas, sources, conversations, and other people’s work—filtered through the writer’s worldview, experience, and own intellectual rigor.
As I have been saying for a while now, the real question isn’t “Did AI write this?”; it’s “Did a human stand inside this work with judgment, care, and responsibility?”
By installing an AI detector, Substack has shifted the burden from authors and readers to algorithms. In doing so, its message to readers is, “Don’t fully trust your own sense of what’s valuable; run a scan and let an AI score mediate your judgment.”
That trains exactly the wrong muscle.
If the main human job in an AI‑saturated world is critical thinking and genuine discernment—the only real defense against slop, misinformation, and the misappropriation of technology—then we need tools and norms that strengthen that judgment, not outsource it to a black box that can’t see thought or care, only patterns.
And speaking of patterns, even those are not a reliable source for AI.
For one thing, when you take into account that many talented, integrity-driven, thoughtful content creators don’t speak the version of language the AI models were trained on, you add yet another barrier to the very human authenticity that we need more of, not less.
Neurodivergent, non-native English, and technical specialist writers are also more likely to be flagged as “AI-generated” simply for writing exactly as themselves.
This is the area where we admit that if English isn’t your first language, or your language at all, you are more than twice as likely to be penalized by an AI detector. (This is my shocked face. 😐)
This is not a bug, it is an entirely wrong premise to be guiding AI from, as the humans leading it.
Also, if you are someone who writes like they talk, but doesn’t necessarily have a librarian’s vocabulary, the “surprise” of your word choices will rank low, even though the content can be both valuable, and human written.
So, if a person is not a writer by nature, or stopped learning the craft of writing past a certain point, it can easily fall into a middle ground area that means nothing against the detector’s parameters, but happens to seem to overlap them.
If “average” is what the detectors are gauging their inspection on, and “average” is where a wrier’s level of writing happens to fall, do we really weed to be collectively adding one more layer of insecurity to an already incredibly hard and vulnerable craft.
I mean, we can’t all be Shakespeare or Dickens.
Best’s real goal, I think, was protecting the promise between writer and reader.
But the tool can’t do the one job it claims. That’s not a bug to patch—it’s a sign the question was wrong from the start.
So, what’s the right one?
🪨 First Principles
There’s No Outside Left to Check Against
AI text detectors didn’t appear out of nowhere; they’re the latest iteration in a longer line of automated text‑analysis and authorship‑attribution tools that evolved alongside increasingly capable language models.
Plagiarism detectors and basic authorship tools compared submitted text against large databases of existing documents (web pages, academic papers, student submissions) to find overlaps and copied passages. And they were fairly accurate at finding direct copying or lightly paraphrased text from known sources, but very poor at detecting original writing or text that didn’t exist in their databases.
And that particular flaw brings us to where we find ourselves now.
Those earlier tools were designed to detect source reuse, not generation by a model. And when LLMs started producing novel text that wasn’t copied from anywhere, plagiarism tools often saw it as “original” and thus not suspicious.
LLM-era detectors invert this supposition: they check resemblance to the model’s own training data. Which makes the detector both the suspect and the judge.
Why?
Because the training data is us—the entire corpus of human writing.
Hang on to that line because it will be important in The Human Layer, too.
So the external reference point that is supposed to keep the tool honest no longer exists. Verification has collapsed into a mirror of itself, creating a self-referential loop—forever serving up its own tail as evidence of its head.
The AI detectors based on the modern LLMs mirror the training data and we are the training data. They also are powerful pattern recognition machines. So, when it scans your writing, it compares you to the average pattern of all of its data—to the statistical average of everyone who came before you.
In effect, reducing your work to the average baseline of humanity.
That’s not comparing you to any fact, and it’s not a coincidental flaw—it’s the direct, structural consequence of detection losing its outside reference point.
We are all flattened out to occupy the mediocre middle where nothing interesting ever happens and nothing insightful, beautiful, or life-changing was ever produced.
As if that wasn’t bad enough, the math gets harder from here. If there is no outside reference—a genuinely unbiased database to reflect originality against—then knowing the origin was never going to be the right question.
The only coherent measurement is responsible synthesis vs. careless collage.
How’s that, you ask?
Synthesis is the deliberate transformation of prior material becoming something new through judgment.
Imagine a chef in the kitchen. They can take yesterday’s roast chicken, a handful of vegetables, and some stale bread and turn them into a completely new dish. The ingredients came from somewhere else; the meal exists because the chef knew what to keep, what to cut, what to combine, and what absolutely did not belong in the pan.
That’s synthesis.
Dumping the entire refrigerator contents into a blender is also recombination. That’s a science experiment with no judgment.
And then there’s a collage: a fragmented copy‑pasted from everywhere and nowhere, placed next to each other without being examined or integrated.
So now imagine someone “making dinner” by putting half a taco, two spoonfuls of cereal, a slice of lasagna, and somebody else’s potato salad on the same plate.
Everything is technically assembled. Nothing, however, has been considered. That’s collage: proximity pretending to be coherence.
And you know it just by looking at it.
Quality—what it is, and how we decide it— was always the actual question, and origin never changes whether something delivers on that.
As much as it pains me that we live in a society that operates predominantly on algorithms, we do. And that fact requires that I also address the governance of it in this equation. Algorithms already seemingly decide so much of what gets traction and visibility—that part is real.
“Yes, and…”
Visibility and value are two different jobs, and in this entire debate, rhetoric in both camps mistakes one for the other.
The algorithm is both overlord and tenant, in my estimation. As long as we remain in the equation, we create the environment for what the algorithm brings to the forefront, what gets visibility.
In a certain way, though, that is how we vote now. We decide what content gets traction and what doesn’t because we as a society, as a whole collective, decide what is valuable.
Only humans, after all, can decide what is good, valuable, and worthy of our attention.
When a reader doesn’t get value from the content in front of them, the move has always been simple—leave. That’s discernment, in real time, doing exactly the job it’s meant to and has always done—telling synthesis from collage.
All without needing a score from a skewed machine suffering from a quasi-Oedipal complex to do it for them.
If the verdict was never the machine’s to render, and never could be, what does it cost that we keep asking it to render one anyway?
🫀 The Human Layer
The Judgment We Can’t Afford To Keep Outsourcing
Though there has been extensive study on the topic, no objective “quality meter” exists.
Humans do, however, have a shared, evolved architecture for judging quality and value—but it’s not a clean, standardized benchmark. We’re human, so it’s a messy, context‑sensitive system that integrates attention, social cues, moral impressions, and personal goals.
That’s precisely why no algorithm can replace human judgment here: the “calculation” of worth is not just about the content; it’s about the relationship between a reader, a creator, and a moment in time.
The Human Calculus Of Worth
Decades of research in psychology, neuroscience, and consumer behavior points to the same conclusion: humans construct value in the act of attending, they don’t discover it pre‑written in the text.
That construction follows patterns.
What we pay attention to gets weighted more heavily in our overall judgment, so attention is not just a spotlight on value but part of the value calculation itself. We rely on extrinsic cues—author reputation, platform, design, social proof—to infer quality, often before we engage deeply with the substance.
And we care about character as much as outcome: trait impressions like honesty, curiosity, and care also influence how we evaluate a piece of writing in ways that raw usefulness cannot capture.
But these shared mechanisms do not add up to a universal scale.
Value judgments remain subjective, filtered through personal beliefs, experiences, and cultural norms. A single paragraph can feel profound to one reader and trivial to another, not because its “actual quality” changed, but because the relationship between reader, text, and moment did.
That’s why only humans can decide what is worthy of their attention.
An algorithm can estimate how many words a model likely wrote, but it cannot replicate the human calculus that integrates attention, social inference, moral intuition, and personal purpose into a single sense of worth.
Substack’s detector treats “AI‑generated” as the key boundary, when the real question readers are answering—consciously or not—is whether a human stood inside the work with judgment and care.
In a world that is actively debating the quality of everything—because AI now touches almost everything, cultivating that capacity and building the muscle for discernment and the calculus of worth is not a nice‑to‑have skill; it’s now the main human job.
There is a growing concern, as well as a growing body of research, around the effects on us stemming from the overreliance on AI. This overreliance measurably erodes confidence in independent reasoning and ownership of your own thoughts and ideas. Essentially the concern is that people are starting to treat AI outputs as final judgment instead of as inputs to their own human judgment.
If the calculus of worth is the muscle, the erosion isn’t hypothetical.
The Arc of Judgment
And as we discussed in the last issue of TEH, to complicate matters there’s a very real adoption-speed dissonance around the whole problem. 71% say AI is moving too fast, a minority trust it to decide anything—and we are adopting it faster than any technology before it anyway. Faster than computers, faster than the internet, faster than even social media, we are relying on a tool that the majority of people in this country don’t trust but still use every day.
In The Event Horizon #001, I asked you if oversight was collapsing judgment. And last issue, if you trust your gut over AI’s read. In this issue I’m asking who’s rendering the value verdict now, and whether we’re noticing that we’ve begun handing it over to the machine we need it most for.
For some time now, the capability has been around to input someone’s work into an AI detector and have it render an incomplete—and often incorrect—verdict. So that story really isn’t the debate at hand.
One of the ways that we exercise our human-only gifts is in deciding what’s valuable, deciding what is good and interesting and thought provoking and entertaining and inspiring and educational, …and what isn’t.
It’s always been that way.
It was that way long before these tools came into being, let alone started being used regularly. And as long as we remain cognizant and intentional, our human discernment will be the reason we stay in the lead of these machines, rather than be led by them.
🌅 Light on the Horizon
Synthesis, Out in the Open
Revoice
Cambridge University researchers built a wearable device, worn at the neck, that translates silent throat-muscle signals into fluent, emotionally expressive speech for stroke survivors living with dysarthria. Early trials show a 4.2% word-error rate and a 55% increase in how satisfied patients feel with their own ability to communicate. This is AI restoring someone’s voice, literally, rather than scoring whether it sounds human enough — the mirror image of everything this issue has been arguing against. Source
The Tree-Mapping Tool
USC researchers built a free tool using aerial imagery and AI to help cities find exactly where urban tree cover runs sparsest, giving planners an affordable way to target shade, cooling, and climate investment in the neighborhoods that need it most. An unglamorous, civic-scale example of AI doing exactly the repetitive, data-heavy work it’s suited for, so human judgment — where to actually plant, who to prioritize — stays with the people who understand the community. Feels right at home in a letter written for a community named after a grove. Source
Computer for Counsel
Perplexity built a legal-research AI where every answer has to trace back to a source a person can independently check, rather than asserting something unverifiable and hoping it holds up. It hands readers the receipts and lets them verify for themselves—the version of AI-assisted judgment this issue has been arguing for the whole way through. Source
🎯 Your Move
This week, run the calculus twice.
Before you publish something:
Could you explain, out loud, to a real person, why you chose this sentence over every other one available to you, and what you cut to get here?
And before you decide something you’re reading isn’t worth your time: is this my own judgment talking, or a score I haven’t checked against it first?
📡 Signal Back
This week’s question:
Where have you caught yourself outsourcing a value call—deciding something is good, worth your attention, worth your trust—to a score, a scan, or a system, instead of making the call yourself?
Tell me where you noticed it, and what you did next.
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.
💡The right idea finds the right person at the right time. You might be the one who gets it there.
💌 A Note on How I Used AI to Write This Issue
Usually this is an extra section in The Understory: a brief, honest note on how I use AI to write the issue. This is a standing section in every The Event Horizon Understory, because my position on AI is simple and steady: use the tools, stay in the loop—stay in the lead.
But for this specific issue, and this particular argument, I decided its transparency needed to be above the threshold this week.
This issue almost didn’t survive its own thesis.
I sat down to argue that judgment can’t be outsourced to a score, then spent the better part of a week working the argument through with my favorite AI tools.
I started with voice notes I transcribed—a rant on what I thought about this whole debate.
Then, after I pulled the individual ideas out of those notes, I took them to Perplexity to research 26 different aspects of my argument, find and read sources, reflect on it all, and ask another query based on the insight it gave me. Back and forth for 26 full turns.
During this process, I also turned to my Claude Cowork. It helped me map out article sections and beats, pressure-test the throughline, and even catch a factual error in my own research brief (Perplexity’s tool is Computer for Counsel, not Computer for Council—small, but this issue argues accuracy matters, so it mattered here too).
None of that structure wrote the argument for me.
The category error, the synthesis-versus-collage distinction, the decision to build this issue around discernment instead of detection—that’s all mine, argued out loud– literally and figuratively, revised more than once because an early draft wasn’t precisely what I was trying to say.
What these tools give me is a second set of eyes fast enough to keep up with how quickly I think—even when I am tripping over my thoughts to get them lined up, a way to do research that would have otherwise easily taken me a couple weeks so I could zero in on what I specifically wanted to shine light on, and a place to catch my own mistakes before you have to.
That’s the whole case this issue is making, really: stay the one who’s accountable for what the tool hands back, and what to do with it, every time.
🌿 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 practice of proving, to yourself first, that your own judgment is still doing the work I just spent this issue arguing only it can do.
This week’s Understory is two things. A Practice Protocol: “The Rejection Ledger“—a simple, ongoing habit of logging every AI-generated line, idea, or draft you deliberately cut, and why, so the discipline behind your discernment has a record even when nobody’s asking to see it. And a Research Deep Dive—the full annotated source set behind this issue’s argument: the bias data, the detector-evolution research, and the studies on what AI overreliance actually costs your independent reasoning, for anyone who wants the receipts behind the case.
Let’s get to work.
For everyone reading from The Canopy: this is what The Understory looks like. Join us in the deeper work.
