The Proxy Collapse

What Happens to People When the Proof Stops Working

Issue 278, August 20, 2026

In the last days of July, a debut crime novelist named Jerry Falade lost just about everything a writer can win. His first book had sold in a fourteen-publisher auction, a two-book deal reported at $2 million. Film interest was circling. Then his own agents withdrew it.

Not because a detection tool flagged it. His agency said it could not authenticate how the manuscript had gone from an idea to a finished book. That was enough. The book was simply doubted.

Falade denies writing it with AI, says he used an open-source model running locally for research only, and points to time-stamped drafts, notes and dated messages documenting how the book came together. He has also said he believes the scrutiny came for him because he is Black. That claim runs through much of the coverage, I cannot settle it from the outside, and I raise it because leaving it out would be its own kind of editing.

He also ran an experiment worth holding in your mind as you read this issue. He told The Bookseller he fed passages from books published in 2012 and 2017 into detection software, and it classified them as 98 and 99 percent AI-generated. Text written years before these tools existed.

It didn’t matter, of course. Within days, deals that would look too good to be true to any aspiring writer were coming apart.

Whether Falade used AI, we may never know, and that’s exactly the point. He was asked to prove a negative. Once the doubt was loose, there was nothing he could say to it.

That’s the world every writer works in now, and increasingly every professional.

A word before you go further. This one runs long. The topic is tangled, and there’s more of the human factor packed into it than I can do justice to at my usual length, and I’m trying not to leave pieces of it on the floor. I think it’s worth your time. We are all walking into something none of us understands yet, personally, professionally, organizationally and as a society.

This is not an AI newsletter. The subject is, as always, the human factor, but of course the growing role AI is playing in our society is creating a complex case study in real time.

The Scanner Arrives

A week before the Falade story broke, Substack handed every reader a button.

Scan any post on the platform and a company called Pangram returns a percentage. This much human, this much AI. Writers can scan their own drafts, dispute the result, add a statement describing how they work, or switch detection off entirely.

Credit where it’s due to Substack. As these systems go, it’s a thoughtful design. It favors disclosure over punishment, and it quietly concedes something much of the industry still won’t say out loud, which is that AI involvement is now a given across nearly every sector of the economy, publishing included, and that it runs on a spectrum rather than a yes or a no.

But one omission shapes everything. Nowhere does the reader see the tool’s error rate.

A percentage looks like a fact, and there’s no way around that perception. It is a guess dressed up as one, produced by tools that are nowhere near as good as the number makes them look. I’ll come back to exactly how far off they can be.

I turned the scanner off for my own publication, and I’m well aware of how that looks.

So let me say what regular readers already know, because I’ve written about it at length in the three-part Artificial Understanding series this spring, on the comprehension gap, the human cost of it, and what feeds the machine (Issues 260 to 262). I work with AI in producing this newsletter. It’s my research assistant and my production support. It is an efficiency that has become very helpful.

This newsletter has never been made alone. For years, it was made with an editorial partner whose voice is still in my head for nearly everything I write, whether that’s an email, a report, a proposal or this newsletter. I learned that much from her. My only regret is that I didn’t have her as an editor earlier in my career, despite my own confidence in my writing, and despite how often other people asked me to write for them.

Today AI sits in that same chair. And here’s the funny thing.

The same things she would delete, my AI editor deletes. A passage that’s off voice. An analogy that only connected inside my own head. The informal word I dropped in to loosen up a serious subject. Same places, same reasons.

I used to push back on her. She would explain. I would lose what I thought of at the time as a battle. I have writing blind spots, and she helped me enormously with them. Now I mount the same battle with a machine, and yes, I still lose. Then and now, for all the right reasons.

What has matured in this phase, and you may catch it week to week, is more formality, more research, and a denser exploration than I used to write. Hold onto that. It comes back later in this issue in a way I did not expect when I started writing it.

The archive behind this newsletter runs to 277 issues. Two hundred and forty of them were written before I brought AI anywhere near this work, which I did not do until November of 2025. I had the same hesitancies most people did, and I held onto them longer than most. My books predate the tools entirely. The ideas, the positions, the arguments, the insistence that something needs saying, the judgment and the final word are always mine, and always have been.

What I declined wasn’t transparency. I’ve offered more of that than any scanner will ever produce. What I declined was being held up against something that can’t tell the difference between a writer who thinks with a machine, a writer who simply writes formally and technically by training and habit, and a machine with no writer behind it at all.

The rest of this issue is about why that difference is the whole question. Not just for publishers, not just for authors, not just for writers. For anyone, in any capacity, personally and professionally.

The Marks Are Coming Everywhere

Substack’s scanner turned out to be the opening act.

On August 11, Anthropic announced that its Claude models will embed an invisible watermark in the text they generate. The idea is clever, and we have been here before with watermarks, tags, files locked against editing, and the rest of it.

The mark doesn’t live in the file, where copying and pasting would strip it out. It lives in the words themselves, in a hidden pattern woven through the model’s choices among equally good phrasings. Copy it, paste it, change the format, and the mark travels along. Rewrite it heavily, and the mark dissolves, because it only exists in the words that survive intact. Break the phrasing, and you break the pattern.

Anthropic concedes the limits, and states them more plainly than I expected. The absence of a mark doesn’t mean a human wrote it, because editing removes the signal. And a mark doesn’t mean a machine authored it. What the mark establishes is that text was processed by Claude, which is not the same claim at all. The model may have only helped edit.

Touched by a machine and thought by a machine are different things, of course. A watermark can only ever speak to the first.

What it can’t speak to, at least not yet, is the consequence. The label, the perception that follows the label, what a reader does with it. What a superior may do with it, or what a superior may think. In these early days it seems premature, and closer to reflex than to sound judgment.

And what strikes me is how much of this is driven by discomfort, anxiety and the identity threat I write about so often. Even if the hype never fully proves out, AI will have significant impact for one reason alone. It put capability in the hands of people who did not have it before.

The direction of this path runs one way, and it has already arrived. The EU’s AI Act rules on this took effect on August 2 of this year. Anything a model produces has to carry a mark a computer can read. Systems already on the market have until December 2 to comply. Anthropic’s announcement came nine days after that deadline landed, which tells you what it was responding to. China got there first, and its labeling rules took effect last September. LinkedIn is testing reader-facing detection. Marks are being built on three continents, with deadlines and budgets behind them.

What is being built nowhere, with no deadline and no budget, is any shared understanding of what a mark actually means. What it should mean. What questions it ought to raise. Whether it means anything at all in our race to leverage AI for all things.

The technology moves, and the literacy doesn’t, and the gap keeps widening. I have written about that pattern more times than I can count, and I watch it happen in person every semester. Students arrive not knowing they live inside a filter bubble. They trust search results with no sense of how ranking decided what they saw. Over ten years of teaching, I have watched that awareness build across a few weeks, and every so often a student walks out of it wanting to go tell everybody else. That is what the gap looks like when somebody closes it.

Now many people trust what AI tells them without question. With our dwindling attention spans, the consequence of that is simply scary.

So, hold onto that gap I mentioned. It’s where the real damage, and the collateral damage, gets done.

The Proxy Collapse

For most of history, good writing was reliable evidence of thinking, and for a simple reason. Producing it was hard work. I can attest to that. Edit after edit. Confirming sources. Making sure a quote is used the way its author meant it. Matching the language to whoever was reading it, because legal does not talk like procurement and neither of them talks like engineering. Checking that the research and references underneath would hold up.

A clear, sustained, well-sourced piece of thinking could only come from a mind that had done that work. Readers never had to check this consciously. The difficulty was the proof. Publishers and universities built their gatekeeping on it, through peer review and official journals and portals. Writers built something more personal on it. An identity, a readership, a following, and most importantly, trust.

Generative AI didn’t break the thinking. It broke the proof. Fluent text can now be produced at almost no cost by anyone, with any amount of thought behind it, including none at all.

That last part is the scariest of them all.

I call this the Proxy Collapse, and I want to be precise about what collapsed. Original judgment, tested expertise and accountable analysis are as scarce and as valuable as they have ever been, arguably more so as the flood of fluent text keeps rising. I raised a version of this years ago, as these tools were climbing, when I asked whether critical thinking was at risk of extinction.

What collapsed isn’t the thinking. It’s our ability to judge the thinking from the surface of the text.

The written word no longer certifies anything.

Seen that way, the detection industry comes into focus. These tools are an attempt to restore the old proof by force. To restore what was comfortable, what was understood, what was valued. Certify that the written word was made by hand and we can go back to trusting it as evidence of thought.

I understand the impulse. It won’t work. Hand-typed text is not evidence of thinking. Machine-assisted text is not evidence of its absence. The thing readers actually care about was never in the keystrokes.

Longtime readers will recognize the shape of this from the Measuring What Matters series. When the real thing is hard to measure, we measure a stand-in. Then we completely forget it’s a stand-in.

And then something worse. The stand-in becomes what people know, and what they base their thinking on. Whether it was ever right stops mattering, because by then it is simply accepted as the way things are. Norms get set and stop being questioned. Think of how often a child asks a parent why, and the answer comes back as well, that is just the way it is.

We Have Been Here Before

Before we accept that assisted writing is some new deception requiring new policing, we should be honest about the practices we already have, and have had for a century, though I would suggest they have existed as long as we have had spoken and written communication.

There has always been something one could call deception in these arrangements. There has also always been an accepted set of practices that nobody gave a second thought to, and probably never did.

Carolyn Keene, author of the Nancy Drew mysteries that three generations of readers grew up on, never existed. A publishing syndicate invented her in 1929, outlined the plots, and hired writers to produce the books. Mildred Wirt Benson wrote twenty-three of the first thirty. She was paid $125 for the first manuscript, under a contract that handed over every right to the text and to the name and required her silence. Her role stayed hidden until she testified in a court case in 1980.

When the novelist V.C. Andrews died in 1986, her estate hired a ghostwriter, Andrew Neiderman, to keep producing books under her name. He has been writing them ever since, for close to forty years now.

I read a lot of those books. They passed around my extended family and became the subject of conversation whenever we got together. I just assumed V.C. Andrews wrote them. I didn’t even know she had died. Who the author was didn’t seem to matter nearly as much as the complicated stories those books told and the shared experience of reading them with family members.

Below the level of books, the practice is so ordinary that we stopped seeing it. The executive’s op-ed drafted by an agency. The CEO’s letter assembled by staff. The leadership memo polished by a chief of staff. We have done that work for clients over the years, and you can add press releases, scripts and speeches to the list.

Nobody scanned those. The industry didn’t call it fraud then and doesn’t now. It was understood as collaboration and as putting honed expertise to work.

I call this the Ghostwriter’s Exemption, and its unwritten terms are simple. Writing by proxy is acceptable as long as it stays expensive and invisible.

What AI did was hand a version of that ghostwriter to anyone with a monthly subscription. The norm didn’t change. Access did. And some of the alarm right now, including the sudden appetite for detection, is the discomfort of a gatekeeping economy watching its private privilege go retail.

One distinction between the human and the machine is worth keeping in front of you. A human ghostwriter exercises judgment and can refuse to make things up when they have the editorial standing to do so. A language model will invent a source with exactly the same fluency it uses to report a real one, which is why machine assistance demands more verification from the named author, not less.

But notice where that lands. It puts the full weight of trust on the named human’s diligence, which is exactly where it has quietly rested all along. Nobody who admired a ghostwritten memoir believed the celebrity typed it. They believed the celebrity stood behind it. They understood that somebody had helped tell the story, and they lost themselves in it without a second thought.

The typing was never the trust. The name was.

There is one more thing sitting under all of this that I’m not going to resolve here, because it deserves its own issue and I intend to give it one. A model can hand you an invented source. It can also hand you words that belong to a real person, or an idea with its author quietly removed, and you would have no way of knowing. A detection score will catch none of the three.

What the Tools Actually Measure

To understand who these tools misjudge, you need to know how they work, and the plain version is this. They don’t know where your words came from. They guess, based on patterns. Text that’s smooth, conventional, and even-keeled resembles machine output. Text that’s uneven, surprising, and idiosyncratic reads as human.

Now think about what that means for careful writing.

Citations follow standard formats. Phrases like “the study found” are formulaic because they have to be. The measured tone of evidence-based work deliberately files off the quirks these tools read as human. And the models were trained heavily on exactly this kind of writing, so executing it well makes your work resemble theirs.

The result is backwards. Disciplined, sourced writing scores as machine-like. Loose, unsupported opinion scores as human. The tools systematically penalize the writers doing the most careful work.

This isn’t theory. OpenAI, which understands its own models better than anyone, pulled its AI text classifier in July 2023, seven months after launching it, citing low accuracy. In its own disclosed testing, the tool caught 26 percent of AI-written text while labeling human writing as AI 9 percent of the time. A Stanford analysis published that same year ran ninety-one English proficiency exam essays, all written by students for whom English is a second language, through seven detectors. On average, the detectors called 61 percent of them AI-generated. Almost every essay in the batch was flagged by at least one. Also that year, Vanderbilt turned off Turnitin’s AI detector entirely. Its reasoning was simple arithmetic. Vanderbilt had run 75,000 papers through Turnitin the previous year. At the 1 percent error rate the vendor itself claimed, roughly 750 of those students could have been wrongly flagged.

Read that again. The people most likely to be flagged for not writing their own work are the ones writing in a second language.

I know this territory from both sides.

As an adjunct professor, I’ve used Turnitin on student work, and I know its usefulness and its blind spots. A decade taught me one rule I’ve never regretted: the tool informs the judgment, it should never be the verdict, and the diligence is still yours to do.

You could fairly ask how that sits with turning the scanner off on my own work, and it deserves a straight answer. A student is being assessed on whether they can do the thinking. That is the entire point of the assignment. You are not assessing me. You are deciding whether to trust me, and you have an archive, a podcast and two books to decide from. Different questions, different tools. What I will not do in either case is put a number in front of a person and call it a verdict.

And I’ve had this happen to me. Some months back I published an issue of this newsletter with Grammarly running in the browser, and its new AI detector announced that entire sections were AI-written. I had written every one of those words myself. There was no appeal. Just an immature tool’s verdict on one side and my own memory of writing the piece a few days earlier on the other.

To be fair, Pangram is better than the earlier generation. It claims it wrongly flags human writing about once in every ten thousand times, and it admits its accuracy drops on short, formulaic text. Stay on that last part for a second, because formulaic text is very often academic and business writing. The same writing my two editors have slowly trained me to produce.

That one in ten thousand is the company’s own number, from the company’s own testing, and researchers have already published cases where it flagged human writing. Even if you accept it completely, the math at scale doesn’t change. Across millions of posts and millions of scans, a steady population of honest writers gets wrongly flagged, each one facing the same impossible task Falade faced, in front of an audience shown a number with no uncertainty attached to it.

It’s worth being clear about what these tools reward, because once you see it you can’t unsee it.

The first generation worked on statistics. How predictable is each word given the ones before it, and how much does that predictability vary across a document. Formal writing systematically strips out both kinds of variation.

Predictable vocabulary. Professional writing leans on precise terminology, clear transitions, and established phrasing. Those pairings are logical and common, so a detector reads them as highly predictable and therefore machine-like.

Uniform rhythm. Formal prose favors well-constructed, medium-to-long sentences with parallel grammar. You rarely swing from a thirty-word analytical sentence to a three-word fragment. That balanced pacing registers as low variation.

No noise. Good writing strips out the tangents, the emotional outbursts, the slang, the typos, the rambling. Detectors treat those exact flaws as the proof that a person was there.

Read that list again and notice what it is actually describing. It’s describing an editor’s job.

I should be fair about where the technology has gone since. The newer tools, Pangram included, threw that statistical method out and trained systems to recognize machine text directly instead. Pangram has published its own argument for why the old approach fails. But every piece of evidence I laid out above comes from the years when the old method was the method. And the newer tools are answering the same question the old ones asked. Only the machinery underneath it changed.

I have worked on this issue since mid-July, for more hours than I can count and through twenty-five drafts, using AI and my own fingers for the research. I keep running aground on the same thing. What these tools call unevenness is precisely what I was taught to remove, so much so that as sentences form in my head, I stop myself from writing the slang, the rambling and the tangents.

The analogies I like. The informal word I drop in to loosen up a serious subject. The tangent that entertained me and would have entertained nobody else. My editorial partner cut those for years. My AI editor cuts them now, in the same places. Both of them were right.

So the machine reads my writing as machine-like because I was trained, patiently and correctly, by people who made it that way.

Funny thing. The quirks that would certify me as human are the ones I have spent a career learning to delete.

The Removal Industry

There’s another half to the detector’s predicament. These tools are losing an arms race, and this isn’t a prediction of mine. It’s already underway, and has been for a while.

An industry of “humanizer” tools has been operating since 2023, rewriting machine text to put back the unevenness that classifiers read as human. Research published this June in the International Journal for Educational Integrity measured the outcome across forty fully AI-generated papers and forty humanized ones. Turnitin scored every fully AI-generated paper below the study’s strict threshold, catching none of them. GPTZero caught one humanized paper out of forty. Pangram held at 92.5 percent on the humanized set, which is presumably a large part of why Substack hired it.

But the rewriting tools will now be tuned against the last detector standing, because that is what always happens next.

The same ready-made industry greeted Anthropic’s watermark. Within a day of the announcement, before Anthropic had even published a way to check for the mark, watermark-remover sites and bypass guides were live.

Sit with what was actually for sale that week. Removal of a mark nobody can yet detect, sold by vendors who can’t prove their product works, to buyers who can’t verify either. A fear market, open for business before the thing it feared had even been documented and put into practice.

Yikes.

I’ve watched this movie for thirty years in digital, and the ending doesn’t change. Digital rights management didn’t end music piracy. It lost outright, while a public ravenous for digital music on brand new devices did what it was always going to do. What ended piracy at scale was Spotify, Apple Music, Amazon Music, and the rest making the honest path easier and far more valuable than the dishonest one.

That’s the lesson the detection era keeps refusing to learn, at least so far. Punish the mark, and you industrialize the scrubbing. Make the mark unremarkable and the workaround economy never grows, because workarounds only spread when the honest path is expensive.

If you want one indicator to watch, watch the removal market. If it stays a fringe of relabeled paraphrasing tools, the honest path is winning. If it attracts real funding and advertises openly, the design has failed, and the fear has won.

The Platform’s Problem Is Real

It would be easy, and wrong, to treat all of this as an unforced error by Substack. The platform has a real problem.

Since AI became widely available, Substack has watched a share of new newsletters arrive that appear to be entirely machine-generated. The platform is a revenue generator for some of its authors (2040 Digital publishes to Substack but not to generate revenue), and a subscription business rests completely on the reader’s belief that a person is on the other end. Readers who find out that a voice they paid for was automated feel deceived, and as I wrote in The Empathy Outsource (Issue 270), that feeling runs backward through everything the voice ever sent them.

Pangram’s own measurements make the scale concrete. Across a million social posts sampled this spring and summer, more than 40 percent of the longer LinkedIn posts came back as fully AI-generated. Longer articles on X came to almost half once AI-assisted writing was counted. Substack, the lowest of the platforms measured, still ran over a fifth once you count both. Worth saying out loud that these numbers come from the same company Substack hired, measuring the problem its product is sold to fix.

Now let me argue against myself, because I am in that data.

My social posts are machine-generated, and while I often edit them, I won’t pretend the drafting is mine. What I will say is what they are. Every one is distilled from an issue I wrote or an episode I recorded. They carry the points of something I actually wrote to another place people are looking. They are not posts with nobody behind them. It saves me time, it keeps the posts on point with what I actually want to say, and it helps me get the newsletter and the podcast in front of people who read, listen or watch.

Somewhere past that line sit the small companies now handing AI their entire social presence, their marketing and parts of their sales, with no source work underneath any of it. The distance between extending your own work and outsourcing your voice entirely is one of the truly unsettled questions of this moment.

The scanner is blind to it. My derived posts and a fully automated feed flag identically.

Substack’s own chief executive, Chris Best, conceded the deeper limit the day before launch. Pangram, he wrote, can only detect whether AI was used to make the text, not whether great human care went into creating it. Those are different claims, and only the second one has ever mattered to a reader.

The problem is very real. The tool counts the wrong thing.

And we do what people always do when something is wrong. We need to act, to put something in place, to be seen doing something about it. The subject I keep coming back to, year after year, is what that impulse costs us later. We are almost always blind to the consequences we did not intend.

The Question That Matters

In June, in The Empathy Outsource (Issue 270), I argued that for personal messages- the condolence note, the recognition, the hard piece of feedback- the authorship is the message. The whole point is that a specific person spent something of themselves on you, and the research shows the value collapses the moment the author becomes uncertain.

Analytical writing runs on a different currency.

When you read this newsletter, or any analysis, you aren’t really asking whether my fingers produced each keystroke. You’re asking whether what I put in front of you is worth your time. Whether it speaks to real issues and real challenges in transformation and organizational management. Whether the sources are real. Whether the judgment behind the argument has been tested and stays consistent. Whether someone stands behind it who will own the errors when they come.

In personal writing, authorship is the message. In analytical writing, accountability is.

A detection score measures the first and cannot see the second. By the scanner’s math, a scholar who spent thirty years building judgment and drafted with a machine scores exactly like a content farm that spent thirty seconds. Meanwhile, a fabricated statistic, typed slowly by hand, scores as perfectly human.

The number isn’t wrong, exactly. It’s answering a question that was never the important one.

The Fear Underneath

I watched what this moment is doing to people at close range this summer.

I’ve written twice now about my session at this year’s Bridge Conference and the conversations that followed it, in Transformation Theater and The Identity Renegotiation. There’s one detail from that room I haven’t put on the page until now.

When the subject turned to AI, and the subject always turns to AI now, I watched the faces change. Not skepticism. I know skepticism well and I can work with it. This was fear. The quickly composed kind, smoothed over within a second because showing it feels unprofessional. It’s the face of someone calculating, in real time, what a new reality means for their own worth.

I did my best with it in the room, and the room met me halfway. People began sharing what had already happened to them and how little idea they had of how to navigate it.

Here’s what makes the fear rational and the scanning ritual strange at the same time.

AI has become part of the basic equipment of office work. Employers aren’t just allowing it; many are actively pushing adoption. Gallup’s most recent numbers, published in July, put AI use at 52 percent of American workers, up six points in a single quarter, with about two-thirds among people whose jobs can be done from a desk anywhere. It has climbed every quarter they have measured it. And the study I mentioned earlier ran more than eleven hundred master’s theses through a detector. Nearly half got flagged. Among those, the middle case showed about 30 percent AI involvement. Help, in other words, not replacement.

Mixed work, some of it yours and some of it the machine’s, is now the ordinary case. A test that nearly everyone fails doesn’t separate anybody from anybody. It just reports that the water is wet.

So why does the appetite for these tools keep growing precisely as they stop being informative?

Try this with me. Imagine a perfect detector. No mistakes, and a complete record of where every sentence on earth came from. Would the unease go away?

My opinion is that it would sharpen. The honest number would be enormous, and every reader would discover that the colleagues they respect, the newsletters they trust, and their own last three drafts all carry the same signature.

A tool that feels exactly the same to you whether it is right or wrong is not measuring anything. It is a ritual.

Which means the worry was never really about where the words came from. Underneath “was this written by AI” sit two questions no detector can reach, and both of them are deeply human.

The first is about the reader. Am I making a decision, forming a belief, arriving at a conclusion, because a machine put something in front of me and I took it?

The second is about the writer, and it is the one I hear most. Am I still necessary? Is the thing I spent years learning to do still worth doing?

I’m not dismissing either one. The first I come back to at the end. The second is the one I want to sit with here, and I think it’s the most legitimate question of this decade. I write about it, I talk about it with clients, and I talk about it with anyone willing to listen. It hurts precisely because it isn’t irrational.

The displacement is real, and I know it from the inside. The editorial and production services I once paid people for, I now perform with a machine. The honest version of that is more tangled than the headline version. I made some personal and professional decisions. As a result, recurring costs had to come down. Several decisions sat behind the one, and the tool was a factor among them rather than the cause of them. Multiply that across every publisher and every marketing department, and the people who feel threatened are not imagining things.

And I should admit the part that had nothing to do with money, because it is the part that implicates me. I saw freedom in it. Freedom to choose subjects that I am passionate about and nobody would have talked me out of, to run long when something needed the room, to pile on sources. My editorial partner held a construct she believed served readers better, and she was right more often than she was wrong. When she was no longer holding it, I stopped being held.

Which brings me back to what I asked you to hold onto earlier. The formality, the extra research, the density you may have felt building in these issues. That is not a machine writing like me. That is me writing without the person who used to tell me when to stop.

And it is precisely what a detector reads as machine-like.

But look at what we’ve done with the question.

Faced with something too large to hold in our heads, we did what we always do. We swapped an unmeasurable question for a measurable stand-in, and then we started arguing about the stand-in. It offers a way to focus when you are facing ambiguity and uncertainty. Transformation dashboards do this. Engagement scores do this. Authenticity percentages do this.

The Human Factor Method treats resistance as signal, and the signal here is clear. An entire profession is being asked to move its identity from producing prose to exercising judgment, with no map and no timeline. It is uncomfortable, it raises more questions than it answers, and it produces the kind of anxiety that turns into identity threat.

The detector is a checkpoint. Almost a way of slowing how fast things are moving by creating a pause. The people putting it up are being human, in the oldest way there is, at a moment that produces fear.

Nobody Is Teaching Us to Read

If the marks can’t answer the question that matters, what we need alongside them is obvious, and it is also the thing that always gets built last. We need a shared understanding of what they do and don’t mean, held widely enough that a score informs judgment instead of replacing it. In every other domain we call that literacy.

And here is the least optimistic paragraph in this issue, because I’ve spent enough years inside standards and policy bodies to know how this work gets done, and what infuses the end result despite anyone’s best efforts and intent.

It won’t come from Congress. I’ve watched privacy bills die in committee for two decades while Section 230 sits untouched, despite every reason to change it to protect the public. The responsibility falls instead to platforms, vendors and industry groups, and my standards and policy years taught me a lesson I’ve never been able to unlearn. Industry builds what the people paying for it need, and often builds their advantage quietly into the result.

Detection tools serve platforms, so they get built. Provenance standards serve enterprise sales, so they get built. Compliance paperwork serves legal departments, so it gets built. Public understanding of what a mark means serves everyone and pays no one.

Detection has a sponsor. Provenance has a sponsor. Compliance has a sponsor. Literacy does not.

History offers one comfort with a warning tucked inside it. Tool stigmas do fade. But they fade when somebody takes on the teaching, when what good use looks like becomes clear, and when enough people accept it. The calculator stopped being a scandal because schools decided what calculator use meant and then taught it. No institution has picked up that job for AI authorship, and until one does, fear fills the space where understanding should be.

Expect the norm, when it finally settles, to settle the way ghostwriting did. By domain. Unremarkable where writing is just a tool. A stigma where authorship itself is the product, which is exactly the territory where every writer reading this lives.

What This Asks of Us

So what does a writer do, standing in the middle of something that used to work and now doesn’t, with nothing settled in its place?

I can only report my own answer. I turned the scanner off, and in its place I point to the record no detector can produce. An archive that predates the tools. Books that predate the tools. Sources you can check in every issue. And a name that answers for all of it.

For writers, the counsel is short. Let the machine carry the load it’s good at. Keep the judgment and the accountability unambiguously yours. Say so specifically, rather than hoping the question never comes up. And when an idea arrives inside a draft rather than out of your own head, turn the tool around and ask it directly what the prior art is and who has written about this. The burden of verification went up. So did the means.

For readers, the counsel is older still. Interrogate the sources, not the writing style. A bibliography you can check will tell you more about the human behind a text than any statistical reading of its rhythm ever will.

And for leaders, there’s a version of this work your organization needs now. An AI policy tells people what is forbidden. A literacy practice tells them what things mean. What a mark does and doesn’t prove. What honest use looks like. What will never be punished if it’s disclosed. If your people believe a flagged document is evidence against their own worth, that belief, and not the technology, is what your next set of results will measure.

The scanner answers whether a machine touched the text. The question that matters is whether a human stands behind it. Those are not the same question, and no percentage will ever make them the same.

The first has a number. The second has a name.

Mine is on every issue, including this one. And here is what the name actually promises, because it isn’t perfection. I get things wrong. I misread a study now and then, and I carry biases I can’t fully see. The promise is that when you tell me, I will look at it, admit it, correct it, and apologize.

That is a promise a person can make and a percentage cannot. A detector will never issue a correction. It has nothing at stake and no one to answer to.

I have both. And that difference, not the keystroke count, is the only thing that has ever made a byline worth anything.

Related Reading

Artificial Understanding, Parts One through Three (Issues 260 to 262). The comprehension gap. What these systems actually do, what it costs the people using them, and what feeds the machine. The fullest account of how I work with AI and where it fails.

The Empathy Outsource (Issue 270). Why authorship is the message in personal writing, and what happens to trust when the author becomes uncertain.

The Confidence Transfer (Issue 271). What happens to expertise when we trust the machine’s answer over the person’s, which is what a detection score asks readers to do.

Transformation Theater (Issue 275). The Bridge Conference session behind the room I describe in this issue.

The Identity Renegotiation (Issue 276). AI adoption as an identity renegotiation before it is a skills transaction. The fear in this issue is that renegotiation made visible.

Is Critical Thinking at Risk of Extinction? The question I raised as these tools were rising, and the one sitting underneath this entire issue.

Why Transformation Dashboards Lie. The proxy problem in its organizational form. Measuring the stand-in and forgetting it is one. Part of the Measuring What Matters series.

The Human Factor Podcast, Season 2, Episode 019: Structural Silence. On why people conceal rather than disclose, which is the dynamic now playing out with AI use across every workplace.

What I Read

Everything below is checkable. That is rather the point of this issue.

The OpenAI classifier. OpenAI launched an AI text detector in January 2023 and pulled it that July for low accuracy. On its own testing it caught 26 percent of AI text and wrongly flagged human writing 9 percent of the time. Both figures are OpenAI’s, published on its own site.

The second-language study. Liang and colleagues, “GPT detectors are biased against non-native English writers,” in the journal Patterns, July 2023. The 61 percent average and the seven-detector test come from here.

Vanderbilt. “Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector,” Vanderbilt University, August 2023. The 75,000 papers and the arithmetic are theirs.

The Markup. Tara García Mathewson, “AI Detection Tools Falsely Accuse International Students of Cheating,” August 2023.

The humanizer study. Van Vlasselaer, Van Droogenbroeck and Spruyt, “Who wrote this? Evaluating the reliability of AI detection tools in higher education,” International Journal for Educational Integrity, June 2026. Every figure I quote on Turnitin, GPTZero, Pangram and the master’s theses comes from this paper.

The humanizer industry. Lars Daniel, “Students Use ‘AI Humanizer’ Apps To Make ChatGPT Essays Undetectable,” Forbes, October 2025.

Substack and Pangram. TechCrunch and Axios covered the launch on July 22 and 23, 2026. Chris Best’s post “Against Claudefishing” went up July 21. Substack’s support documentation describes how the feature works. The platform percentages come from Pangram’s own study of just over a million social posts, published July 9, 2026. Company research about the problem the company sells against, which is why I said so in the text.

Jerry Falade. The Wall Street Journal broke it, and Publishers Weekly followed, August 2026. Deal figures, the agency’s statement and Falade’s response as reported there. His account of the 2012 and 2017 test comes from The Bookseller.

Anthropic’s watermark. Announced August 11, 2026, and covered by TechCrunch. Anthropic’s own wording is that a watermark only helps test whether Claude might have produced or processed the content, and cannot separate Claude writing something from Claude heavily editing it.

The rules. The EU AI Act’s transparency requirements, in force August 2, 2026, with existing systems given until December 2 to comply. China’s labeling measures were issued in March 2025 and took effect September 1 that year.

How detection works. Gehrmann, Strobelt and Rush, “GLTR,” 2019, and Mitchell and colleagues, “DetectGPT,” 2023, for the statistical approach. GPTZero has said publicly that it stopped using that approach in autumn 2023.

Workplace numbers. Gallup, “Organizational AI Adoption Jumps Six Points,” July 2026, surveying 22,573 employed American adults. The 52 percent and the six-point quarterly jump come from there.

Nancy Drew. The University of Iowa Women’s Archives on Mildred Wirt Benson, and JSTOR Daily on the Stratemeyer Syndicate. Her authorship became public when she testified in the Grosset & Dunlap litigation in 1980.

Join the Conversation

Connect with 2040 Digital on LinkedIn and subscribe on Substack.

The Truth About Transformation

The Truth About Transformation: Leading in the Age of AI, Uncertainty and Human Complexity (2025) explores why transformation fails at a 70% rate and what it takes to lead change as the human problem it actually is.