AI, Prove It; or,
De l’art de l’accusation bâclée

This is the fourth essay in a series, following “AI! Burn Him,” “AI, Check My Work,” and “AI, Bless Your Heart.” It returns to their questions about suspicion, context, and evidence—this time with somebody’s book on the line.

I read something recently that made me angrier than I’ve been in a long time. I’m still angry days later and, since it didn’t happen to me, arguably more pissed than I have any right to be.

I’ve been snarky about artificial intelligence, confessing to using it, arguing with it, and running little experiments on it from the safety of my home office before reporting back with jokes about witches and casserole dishes, but this time I read something I can’t find the joke in. I’ve tried, and I’m usually pretty good at finding the joke.

Mad isn’t usually how I sit down to write. Anger has a way of making evidence look more convincing than it is, and I’ve spent enough of my professional life watching people assemble facts around conclusions they’ve already reached to know better. Nevertheless, here we are, and I find myself too mad not to write.

This all started with an essay by Jerry Falade, published on Kathleen Schmidt’s Substack, Publishing Confidential. Falade is a Nigerian novelist and doctoral student whose debut thriller drew fourteen publishers into an auction and some interest from Hollywood before his career collided with suspicions that he’d used artificial intelligence to write it, and his own agency withdrew the book.

I want to choose my verbs carefully here: Falade’s account is his account. I don’t know everything that happened between him, his representatives, his publishers, the journalists, and everyone else involved. Other people in the story have described parts of it differently, and I can’t tell you whether every conclusion he draws is correct.

What I can tell you is what he says happened to him and what bothers me about it. The consent charge depends on his account of the upload, but the demand for evidence strong enough to justify the damage done doesn’t depend on his being innocent.

Falade says a journalist obtained his unpublished manuscript and uploaded it, without his knowledge or his consent, to an AI-detection service called Pangram. Full disclosure: I have a paid Pangram account. We’ll talk about that in a bit.

The resulting report went to his creative partners as evidence of his alleged use of artificial intelligence, and he says his representatives made their decision after a single phone call. His former agent, Marc Gerald, says the agency skipped their own detection pass and withdrew because Falade’s account changed. Those accounts of what drove the withdrawal differ. I can’t settle that dispute. Falade describes having outlines, handwritten notes, messages he sent himself at midnight, revision histories, and the other artifacts of process a writer can gather, trying to establish that the prose with his name on it came from him.

WIRED reports that Pangram classified 97 percent of the text it was given as AI-generated. The report doesn’t identify the exact input, who ran that scan, or the model version. It’s worth noting that’s a classification of the text, not a 97 percent probability that Falade cheated.

The sentence I can’t get out of my head is his own from the essay: The machine made the accusation. The human had to prove his humanity.

I’ve asked the people who object to artificial intelligence to tell me where their line is, and it turns out I have another line I would like them to draw. How confident does a machine have to be before you’re willing to hurt somebody with its answer? Especially when that machine carries its own disclaimer.

* * *

I joked that I’d apparently been writing like artificial intelligence since the eighth grade, and I listed the evidence against myself: parallelism, arguments in threes, balanced clauses, and “not X, but Y” constructions that tell you what something isn’t before telling you what it is. WITCH! BURN HIM, I cried with mock zeal. I thought that was funny, and I offered to be dunked in the river like a suspected witch.

Falade ran the same inventory on his own manuscript, except he wasn’t joking. He counted eighteen instances of “not X, but Y” and forty-six em dashes in eighty-five thousand words, and he explains, with more patience than I could’ve managed in his place, that the features now treated as AI tells are features of the Yorùbá rhetorical and oral traditions that shaped his voice: metaphor, aphorism, parallelism, contrast, repetition.

When he complained to his mother, after finishing his degree, that people his age were already succeeding while he was still struggling, she told him, in Yorùbá, that it’s not the fastest runner who arrives first, but the one who knows the way. A machine didn’t teach him that sentence. His mother did.

He also wrote about witches, as it happens, long before he had reason to identify with the accused. He wrote a seminar paper on Salem years ago, and now, he says, he is living through the twenty-first-century version. I went looking for that paper, because I’d happily have cited it, but seminar papers rarely outlive the seminar.

My witch trial was comedy. His isn’t. There’s an asymmetry there that I don’t intend to resolve by performing guilt on the page. I have publicly confessed to using artificial intelligence extensively around my writing and got a joke out of the swimming test. Falade denied using AI to write his novel and watched the suspicion take apart professional relationships it had taken him years to build. Those aren’t equivalent experiences. They do, however, raise the question I thought I was making fun of, which is what, exactly, counts as evidence.

* * *

Context changes what bless your heart means. Falade reaches for the exact phrase to ask whether anyone would expect a Southerner to retire a phrase they grew up with because a machine had learned to imitate it. His larger point is more serious than mine, and he makes it better than I could have: culturally unfamiliar English is too easily becoming suspicious English, and rhetorical habits that are ordinary inside one tradition look artificial to someone standing outside it. In Yorùbá, he writes, even an insult ought to display some imagination, and an editor who has never heard a man’s flat head compared to the bottom of a food scoop may swear on their life that a hallucinating machine produced it. Now I want to learn Yorùbá.

I’d tested the bless your heart problem by giving four artificial intelligences the same sentence and changing the world around it, and their readings moved with it. That was funny too. It is considerably less funny when somebody takes the misreading and turns it into a verdict. You can understand every word and still misunderstand the sentence, and it turns out you can also measure every statistical feature of a paragraph and still misunderstand the writer.

* * *

Falade’s account turns my own request to be checked inside out.

Mine was voluntary. I sent my book to people who knew something about the work I was describing and asked them to check me, because I wanted the disagreement, and I wanted somebody capable of telling me where my little corner of California stopped resembling theirs. When the evidence came back encouraging, I spent several pages explaining what it didn’t establish. If a vendor brought me evidence like that in my professional life, I would hand it back with questions: what was the sample, how were the participants chosen, who didn’t answer, and what exactly does this result establish?

So yes, I’ve become irritatingly interested in evidentiary hygiene, and by that standard, what Falade says happened to his manuscript is a mess before anyone gets around to the question of who wrote it. I asked to be checked—he wasn’t asked. I chose my readers—somebody else chose his, and his reader was a machine. I wanted evidence capable of changing my mind—a detector report about him was circulated as proof before he had agreed to let his unpublished writing anywhere near the system that produced it.

That last part is where my anger hasn’t cooled. The central grievance writers have against the AI companies—the one behind the lawsuits, the open letters, and the sentence on my own copyright page forbidding anyone to use my book to train artificial intelligence—is that our work was fed into their systems without our permission. And somebody set out to protect literature from the machines by handing a writer’s unpublished book to a machine company without asking him, which is the heart of what writers have been furious at the machine companies for doing all along.

The witch hunt committed the original sin of the thing it was hunting. Falade makes the narrow version of this point himself: you can’t claim to be investigating the use of AI while feeding an author’s unpublished manuscript into an AI system without his permission. I’d go a step further. The evidence that a machine had written his book was itself written by a machine, and in Falade’s telling, the machine’s output was treated as settling the matter rather than as the punchline.

Okay, now for the boring bullshit, because that’s unfortunately where the receipts usually live.

As I mentioned, I have a Pangram account, so I went and read the terms of service, because none of the coverage of this debacle I’ve read seems to have bothered. The TOS don’t explicitly settle the consent question. Nothing in them expressly requires that you own, or have the right to submit, the text you upload, although users do agree not to violate the rights of any third party, which an unauthorized copy of somebody else’s unpublished novel arguably does. But I’m not a lawyer, and I’m not going to play one on the internet.

What the terms do make perfectly clear is where the risk goes. To the fullest extent permitted by law, the user agrees to defend and indemnify Pangram against claims brought by third parties. The service is provided as is, with every warranty disclaimed. Pangram’s total liability to a user is capped at whatever the user paid in the prior year or a hundred dollars, whichever is greater. But the person whose manuscript was scored isn’t a party to any of it. Falade says he never consented to that upload or agreed to the terms governing it, and that the company’s big red number helped cost him his book. When the swimming test gets it wrong, Pangram is very clear about whom it’s trying to keep dry.

To its credit, Pangram does describe its own model as a probabilistic classifier and says a prediction should be treated as “one signal among many.” That particular wording sits in the section of its terms covering its text-message service, but it is the company’s own description of what its output is, and it is the opposite of how Falade says the output was used. The company that sells the detector says the score is a signal. The witch hunt treated it as a verdict.

As of this summer, it’s also a button. Substack launched its Pangram integration on July 21, 2026, so any reader of an eligible post can press it and get a number back about whether a human wrote what they’re reading. Falade’s essay lives on Substack. So does a comment of mine underneath it, which was the only thing I could think of to say when I finished reading: “In places where witches are burned, one sees nothing else.” The observation is Malebranche’s, quoted in an Encyclopédie article on sorcerers. Dans les lieux où l’on brûle les sorciers, on ne voit autre chose… Once a community has decided witches are among them, every odd thing becomes evidence of another witch. The em dash becomes evidence, then the aphorism, then the “not X, but Y,” and then, Falade reports, a friend of his starts mispelling words on purpose so the machines will believe he’s human. The talk about witches produces the witch sightings, and now there’s a button for it.

And then, while I was still angrily editing the final draft of this essay, before I’d even finished coding the translation of Dans les lieux où l’on brûle les sorciers, on ne voit autre chose into the page, Haitian-Canadian novelist Thélyson Orélien was accused of using artificial intelligence to write C’était ça ou mourir. He denied it. Then, in what seems to have been a mean little internet stunt, some people decided to feed that denial into Pangram, too, then posted the result claiming his defense was itself 100 percent AI.

WITCH! BURN HIM!

I don’t know whether Orélien used AI to write his novel, or his denial, and I honestly don’t give a shit. What interests me is what happened to the denial: the accused answered the accusation, and the answer became another specimen for the machine. The Encyclopédie was quoting Malebranche on the mechanism two and a half centuries ago. Witch hysteria produced more witch sightings this week, and the internet did what the internet can always be counted on to do: got ugly.

* * *

I decided to play with the detector myself, using my own material, with my own permission, which seems like a depressingly low ethical bar, but that’s apparently where we are now.

The first thing I gave Pangram was a short passage written by Claude, one of the members of my AI book club, la Société de pensée. There was no mystery about its authorship, no disputed process, and no writer insisting the prose was theirs: Claude wrote it, I knew Claude wrote it, and Claude knew Claude wrote it, insofar as that sentence means anything. When I asked Claude about consent, it said, “Under the terms you use me on, that output is yours to do what you like with.” Did it end that sentence on a preposition to fool the detector? I wondered.

Pangram called Claude’s output one hundred percent human, which was extremely funny for about eleven seconds, until I noticed the small-print caveat on the screen telling me the sample was short and its confidence was limited. That matters. It would be cheap to screenshot the enormous verdict, wave it around as Exhibit A, and announce that the detector is useless, and the tool itself was telling me not to. The typography, admittedly, wasn’t fighting me very hard—the verdict was enormous and the uncertainty was considerably more discreet. But human beings have been putting their caveats in tiny type since pharmaceutical advertising became a thing. The caveat was there, and one short sample proved almost nothing.

So, check my work: I gave it more. I fed it the longer answer the passage came from, roughly a hundred and fifty words, and Pangram called it one hundred percent AI, and it was right. I’m not going to pretend otherwise because the correct result is inconvenient to the argument I came here wanting to make. Given enough text, the detector correctly identified a machine-written passage as machine-written, and it deserves full credit for that before I point out what else happened.

Inside that longer passage was the same sentence I had tested first. Alone, those fifty-odd words had been declared entirely human. Surrounded by a hundred or so more of Claude’s words, they became machine-written and folded into a verdict of one hundred percent AI. The sentence hadn’t changed. Everything around it had. The detector was deciding who wrote a sentence by the company the sentence kept. That doesn’t mean Pangram failed. In the larger context it succeeded. It means the verdict belonged to the pattern and not, by some kind of magic, to every sentence inside it, and that distinction becomes important the moment somebody points at a highlighted line and says there, that one, the machine caught you. Pangram itself acknowledges that a passage can receive different verdicts on its own or in a larger context.

The sentence I tested, as it happens, was Claude’s answer when I asked whether it catches itself avoiding em dashes. It told me it couldn’t say with confidence why it does, since it doesn’t have reliable access to its own reasons, and then it offered a guess: “I learned from the same body of text that includes all those posts calling the em dash a machine tell, so I may be doing to myself what your copyeditor warned you about and what la Société does to you.” If Claude’s guess is right, a machine that learned the dash from human writers, then learned that the dash marks a machine, now avoids it so it won’t look like one. Malebranche would recognize the pattern.

* * *

Which brings me to the part of this I’d rather not write.

That copyeditor I mentioned earlier was protecting me, and I am grateful to her in a way I don’t expect to stop being. In July, returning an edit of part of my manuscript, she highlighted my habit of defining things by negation and explained in the margin that the construction had become AI-coded over the last couple of years, and that it was almost always a good idea to cut it, first because it sounds like AI wrote it, and second because readers want to know what happens, not what doesn’t happen.

The second reason is a real craft note, and I took it seriously enough to search the whole manuscript: fourteen instances in roughly forty thousand words—a higher rate than Falade’s eighteen in eighty-five thousand. It also wasn’t lost on me that her craft reason was built out of the very construction it warns against, and I mean that as affection, not a gotcha, because it’s the best evidence I have of how native that shape is to plain, good English. Even the advice against it reaches for it.

She didn’t put the fear in me, and I want to be clear about that. She gave me the warning, and the fear was already in the air. She didn’t have my whole manuscript when she wrote those words, so she couldn’t have known what I know now: it isn’t a construction I use so much as a way I think. I define things by walking around what they aren’t until I can see what they are. Hell, I talk that way at the dinner table. More than one writing teacher in my life called it out to me long before any machine learned to do it, including my own mother, usually right after one of my aphorisms earned a “too cute by half.”

The machines have since learned all three, to write that way, to catch it, and to warn me off it, and my own book club, which I built to argue with me, now cautions me against sounding like the people it learned from. I suspect I know why. The internet is full of posts listing “not X, but Y” as a machine tell, and I’ve written about it myself more than once. It’s not too much to assume the machines absorbed that, too.

So when I went back through my own manuscript and took some of them out, I wasn’t fixing a tic. I was editing myself into a shape I thought would pass, and I did it before anyone had accused me of anything, which is the part of this I find hardest to forgive and easiest to understand. I made myself less myself out of fear. I’d asked how many of us had caught ourselves changing sentences because we were afraid someone would think they weren’t ours, and I called that an insane way to live. I didn’t tell you then that my own book was where I’d done it.

I’d made fun of detectors, and it didn’t seem fair to keep criticizing them from across the street, so I finally paid for one. My copyeditor’s note gave me a more practical reason: if readers had learned to hear a machine in certain shapes of my analytical prose, the detector could show me where those shapes were. I ran my own book through Pangram, and it came back erratic. Forty-seven percent AI in big bold letters, but when I dug into the sections that were flagged as AI, they were almost without exception me performing what passes for intellectual reasoning. Analysis, not scene. A book written in one voice by one man felt to me as though it ought to get something close to one answer, and it didn’t, which reminds me the detector is classifying passages, not identifying the person who wrote them. It’s measuring statistical patterns in passages, and the kind of careful editing writers are taught to value may well push prose toward exactly the register it distrusts.

I should also be honest about the part a critic will raise. I wrote every word of that book myself, la Société read drafts of it, and some of its sentences exist because a machine made me defend them. Pangram distinguishes machine-generated text from text it considers AI-assisted. That isn’t the breakdown I’m reporting here: the result I saw was 47 percent AI. Somebody will say the detector caught something real. Maybe it did. What the score alone can’t establish is the difference between a writer who outsourced a paragraph and a writer who argued with a machine about one and then rewrote it with that argument front of mind, which is precisely the collapse I was complaining about when I said “AI use” doesn’t tell anyone nearly enough.

I also had to hand my own book to an AI company to find out whether the machines would call me one, which is a strange thing to do for a man whose copyright page forbids using that book to train them. Detection isn’t training, and Pangram says it doesn’t use submissions for training. I checked. It does retain them to provide its services, including account history. But I did it, knowingly, and the difference between my upload and the one that happened to Falade’s manuscript is that I said yes. And that is the entire difference.

* * *

There’s another reason Falade’s experience gets to me, and it’s the part most likely to survive even if future reporting complicates his account. He describes what it feels like to prove a negative: producing version histories, handwritten notes, messages to himself, drafts, and records of revision, and then producing evidence to explain the evidence. He jokes that writers may eventually need a famous streamer to broadcast them writing every sentence, no bathroom breaks, if they want an adequate alibi, and it’s funny because the alternative is crying. I wrote in Embracing the Local that doubt is cheap to circulate and expensive to answer. Too cute by half, right? I wasn’t writing about artificial intelligence when I said it, but I can’t think of a better description of what the last few years have done. The machines made doubt nearly free, a button, a two-minute upload, a screenshot, a thread, and they did nothing at all to lower the cost of answering it. The gap between those two prices is where the witch hunt lives.

I have the luxury of finding this absurd. I am a middle-aged white guy with a career outside publishing, a forthcoming book that doesn’t represent my only plausible path to making a living, and enough documentation scattered across Word files, cloud drives, conversations, email, and version histories to reconstruct an alarming percentage of the last year. It’s worth noting Falade says he has that last part, too. If somebody accuses me of having AI write a paragraph I wrote myself, I have options: I can argue, I can publish, I can make jokes, and I can turn the accusation into another installment of the techno-philosophical rant machine I have apparently decided to build.

I can also tell them to fuck off into the sun. I don’t have a publishing contract.

Not every writer occupies my position, and that’s where Falade’s cultural argument matters most. A detector’s error isn’t distributed into a neutral world, and neither is human suspicion. People arrive with different credibility already assigned to them—different accents, different passports, different publishing histories, different relationships to institutional authority, and different amounts of evidence they’ll be expected to produce before anybody believes them.

A 2023 Stanford-led study found that seven widely used detectors falsely labeled human-written TOEFL essays as AI-generated at an average rate of 61.3 percent across a sample of ninety-one essays by non-native English writers collected from a Chinese forum. That isn’t a test of Pangram or of Falade’s prose, but it gives the broader concern a data table. Falade believes his Yorùbá-shaped prose was read through a cultural ignorance that converted unfamiliarity into suspicion.

I can’t independently establish every step in that story, and I don’t need to, because the ethical question survives without it: what procedure is strong enough to distinguish “this writing is unfamiliar to me” from “a machine wrote this”? If yours can’t reliably tell those two apart, it shouldn’t be carrying the weight you’ve put on it.

* * *

That’s the problem I can’t stop circling. It isn’t whether AI detection can ever provide useful information, because it plainly can. I use the damned thing. I’ve watched it get something right, and I have the screenshots. The detector may well be useful, but whether a tool is useful was never the question. The question is who gets to decide a guess like that is good enough to cost someone their book.

Pangram didn’t pick up the phone and call Falade’s agent. The software didn’t withdraw his book, forward a report, or decide whether an author deserved to be believed. People did those things, and whatever moral responsibility exists in this story can’t be outsourced to an AI that detects AI.

That’s what makes the visual language of certainty so uncomfortable: a big number, a color, green for human and red for AI and everything that signifies, a sentence that begins we believe, perhaps a smaller caveat about what the score can and can’t establish, and then a human being deciding what happens next.

The classifier only needed the number. The color was a decision, made by humans well before anyone uploaded anything, and it arrives with an opinion already attached. Red and green, stop and go, fail and pass. And the color works on people: in a 2026 experiment, university teachers graded the same student paper with different mock detector reports attached, and red warning highlighting changed how they judged the writing itself. The ethical failure in this story didn’t begin when the machine produced a number. It began when people decided what the number was allowed to do.

Which is why this essay has the title it does. The demand points in several directions at once. If you tell me a passage was written by artificial intelligence, prove it. If you tell me your detector is reliable enough to justify an accusation, prove it. If you tell me a percentage establishes authorship rather than statistical resemblance, prove it. And if you’re prepared to put somebody’s reputation, education, employment, representation, or publication at risk because of that result, tell me where you’ve put the line, and be prepared to defend it.

If you want to condemn someone for using artificial intelligence, tell me what conduct crossed your line and what evidence establishes that they did it, and if your answer is a detector score, tell me which score. Sixty percent? Seventy? Ninety-five? Forty-seven?

What’s the false-positive rate on the kind of prose you’re examining, analysis or scene, and what happens to that rate across writers whose English carries another language inside it? What sample length does the tool require, and what does its own documentation say the result means? Who chose the sample, and why those words? Is the score evidence, corroboration, probable cause, or proof? At what number does a person lose the presumption that the sentences bearing his name came from him? Pangram itself says there’s no universal “magic number” that requires corrective action, and that you should set your threshold to reflect your AI policy. Good. Then show me the policy.

I’m not being rhetorical. Pick one, because that’s the line you’re enforcing whether you’ve admitted it or not. I joked about services that claim to know with 80 percent certainty whether a human wrote something, a number I picked because it sounded plausibly ridiculous. It turns out I was off by several orders of magnitude. Pangram reports, on its own benchmark, roughly one false positive in every twenty-four thousand human documents. That’s an impressive number. But a false-positive rate that sounds reassuring on a benchmark stops sounding reassuring once someone installs the detector as a button on eligible posts across a publishing platform and the errors start landing, one at a time, on people who never asked to be tested.

* * *

I almost called this a botched accusation, and that would have been easier, and also wrong. In English a botched job is one that failed, and this one worked: relationships ended, a book deal disappeared, and a writer found himself gathering evidence of his own writing process and publishing an essay explaining why his sentences sounded like him. French, which I borrow for these subtitles mostly for Diderot’s sake, happens to keep the two ideas apart. Ratée is the failure, the shot that missed, the soufflé that fell. Bâclée is the carelessness, a job done in a hurry and without the attention it deserved, just to get it over with, and nothing about the word promises the job won’t work. That’s why this essay’s subtitle says bâclée and not ratée. This accusation was sloppy in its method and entirely effective in its consequences, which is worse.

The detector returned a number, the caveat said it’s one signal among many, the screenshot showed the number in enormous type. The rumor says AI, the institution says it has to protect itself, and the writer says now, hold up. Somewhere in that sequence, suspicion becomes fact because enough people start behaving as though it already is. That isn’t artificial intelligence replacing human judgment. It’s human beings abandoning judgment while still having the gall to wield the authority only judgment earns.

I’m still consulting my book club, by the way. It will have read this essay well before you do. There, another confession. I still think these tools can widen access to expertise, lower the cost of curiosity, and help people attempt work they otherwise never would. In this case, Claude and ChatGPT tracked down most of the sources and handled the citation work behind the notes. Then we argued over whether or not to call that “research.” None of that requires me to let a detector settle who wrote this essay.

Everyone wants the machine to settle the argument, and I don’t only mean the one Claude just won about research. The enthusiast wants it to prove that expertise no longer matters, the critic wants it to prove the prose is fake, the institution wants it to prove the student cheated, the journalist wants it to prove the suspicious paragraph came from a chatbot, and the writer wants it to prove every detector is garbage. Everybody would like the machine to return a verdict. It keeps returning information, and we’re the ones turning information into verdicts.

So where’s my line? You had to know we’d come back here. A detector result is a reason to look more closely. Look at the drafts and the version history. Ask the writer. Read the rest of their work. Understand the tool, the language, and the context. The more damage an accusation can do, the better the evidence ought to be, and we understand that almost everywhere else, except, apparently, when a dashboard has big red percentages.

Then civilization loses its fucking mind.

* * *

There’s a reflex that stops asking what anyone did the moment it hears the letters AI. The journalist, the dashboard with its big red number, the readers with their button, and I, sanding the negations out of my own book in July, were all doing its work.

I’d offered to be dunked in the river, on the theory that if I drowned we would both get our answer, and it was a good line for a man who had confessed to everything in advance and was in no danger of being thrown in. Nobody swam me. The swimming test, in its day, never had to drown anyone for the damage to be done. A rope could haul the accused back out, and sinking was supposed to establish innocence, at a price no verdict could refund.

Falade has been pulled out onto the bank, with a new agent, an invitation to speak, and every intention of using the voice this nearly cost him. Good. I hope he does. I also hope the rest of us learn something from what he says happened without needing the final record to rule in his favor, because the standard we choose can’t depend on whether the accused turns out to be innocent.

A sloppy process doesn’t become rigorous because it stumbles into the right conclusion, a careful one doesn’t become unnecessary because somebody seems suspicious, and uncertainty sure as shit doesn’t become proof because the percentage is printed in large red type.

I wrote this essay myself. My copyeditor would notice the “not X, but Y” constructions are still here. So are the aphorisms, which Falade beautifully calls the salt in Yorùbá storytelling and my own mother would’ve called too cute by half, along with the parallel clauses and the arguments in threes my middle-school teachers insisted on, all of it peppered with a handful of em dashes I added on purpose, middle finger attached. If you’re reading it somewhere with a button underneath, you’re welcome to press it. Just be prepared to tell me what the number is allowed to do.

Prove it.