This Post Was Written With AI

Scan this, motherfuckers.

Warning: if you are sensitive to profanity, this post is not for you.

Recently, Substack CEO Chris Best announced a rollout of AI scanning integrated into Substack, purportedly to help humans know when something they are reading has been “written by no one”. Hold onto that thought – it’s going to be important later.

There’s a scene in “Team America: World Police” where team member Gary offers the following monologue:

“We’re dicks! We’re reckless, arrogant, stupid dicks! And the Film Actors’ Guild are pussies. And Kim Jong Il is an asshole. Pussies don’t like dicks because pussies get fucked by dicks. But dicks also fuck assholes. Assholes who just want to shit on everything. Pussies may think they can deal with assholes their way, but the only thing that can fuck an asshole is a dick with some balls. The problem with dicks is that sometimes they fuck too much, or fuck when it isn’t appropriate, and it takes a pussy to show ‘em that. But sometimes pussies get so full of shit that they become assholes themselves. Because pussies are only an inch and a half away from assholes. I don’t know much in this crazy, crazy world, but I do know that if you don’t let us fuck this asshole, we are gonna have our dicks and our pussies all covered in shit.”

Now I want you to sit with that, because there are assholes all over this site trying to cover everything in shit, and it’s going to take some dicks willing to fuck them to clean this shit up.

Arvind Narayanan ran the arithmetic on Pangram’s claimed false positive rate of 1 in 10,000, taking the number at face value. Apply it to students submitting 500 to 1,000 written works across four years of college and 5 to 10 percent of the student body gets falsely accused at some point. Now apply it to Substack. Millions of posts, notes, comments, and replies, each scannable by any asshole with a suspicion, and the scans are not random samples; they’re adversarial by construction. Nobody scans a post they trust, and 99% of the people scanning a post in this way are assholes. The population being tested is pre-filtered for texts that already smell wrong to assholes, which means the priors are polluted before the classifier ever runs. A 1-in-10,000 error rate executed millions of times, concentrated on disputed texts, produces a steady stream of human writers holding a screenshot that says they don’t exist, held up by assholes who want to discredit them.

The Pangram benchmark tests pure human text against pure AI text. Almost nobody on Substack is producing either. The realistic case is the writer who outlines by hand, drafts with a model, rewrites half of it, keeps three sentences she couldn’t beat, and edits the rest across four sessions. On this text, the state of the field is blunt: detection accuracy collapses by 25 to 50 percentage points on paraphrased, hybrid, or lightly edited text. A 2025 study, “Almost AI, Almost Human,” found standard detectors misclassify AI-polished text as fully human between 10% and 75% of the time. And for anyone deliberately evading, the adversarial paraphrasing work presented at NeurIPS 2025 showed a training-free attack that drops detection rates by roughly 80 to 88 percent across detectors, including ones adversarially trained to resist exactly this. Pangram knows; their own DAMAGE paper studied 19 commercial humanizer tools and demonstrated that most detectors fail against them. They built defenses. The humanizer vendors read that paper too. This is an arms race where the attacker iterates in minutes and the defender retrains on a schedule, and the economics of that asymmetry do not favor the defender.

The assholes won’t tell you this. They’ll tell you that the high level tests have nearly 100% accuracy against entirely AI generated text, while utterly ignoring what happens downstream: ANY use of AI their scanner detects gets held up to denigrate a writer’s work, wielded by pussies so full of shit that they have become assholes themselves. According to them, if you haven’t written your post into a Moleskin notebook in long form cursive handwriting using an ink-dipped quill, then it is somehow less than the work they have produced.

So map the terrain. The detector is near-perfect on raw, unedited model output: the text nobody was going to be fooled by anyway. It degrades precisely where the interesting cases live, in the enormous middle territory of collaboration, assistance, and revision where most working writers now actually operate. The instrument is sharpest where it’s least needed and dullest where the entire dispute happens. This is exactly the environment that assholes thrive in, and why dicks are sorely needed to fuck them.

Now let’s talk about false negatives. This is the side nobody talks about. A false positive wrongs one writer. A false negative wrongs every reader, because the scan doesn’t return “inconclusive”; it returns a percentage that reads as verification. Run humanized slop through the scanner, collect your low score, and the platform has now laundered your content with an institutional credential it never earned. The 40% figure Best cites for AI-generated text on some platforms came from Pangram’s own estimates; the operators producing that volume are exactly the ones with the incentive and the tooling to defeat the scan. The writers most likely to eat a false flag are the ones doing nothing sophisticated at all: the ESL writer, the technical writer whose prose runs formulaic, the blogger whose voice happens to sit near the distributional mean. The scan punishes the artless and credentials the industrialized. That’s not transparency. That’s an asshole who needs to be fucked by a dick.

Which brings us back to the phrase doing all the rhetorical work. “Written by no one” is a metaphysical claim wearing a percentage as a costume.

Here is what the classifier actually measures: whether the statistical texture of a token sequence resembles the outputs of known models. Here is what readers actually care about: whether somebody meant this. Whether a person chose the argument, stands behind the claims, would be embarrassed by the errors. Those are different questions. They are not even adjacent questions. Best’s own concession that Pangram can’t detect care admits it; a tool that cannot see care cannot see authorship, because care is what authorship IS. The single mother drafting her first post with heavy model assistance because English is her third language has an author. The content farm hand-typing paraphrased slop at volume does not, in any sense that matters, and the scanner ranks them backwards. That’s asshole math.

I’ve spent a year and a half studying what I call machine pareidolia: the way humans pattern-match minds onto model outputs, and the way models pattern-match mind-shaped text back at us. Neither side of that exchange can verify what’s behind the pattern. Substack has now productized the inverse operation. The scanner looks at statistical texture and certifies the absence of a person, no-one-ness as a service, with the same unearned confidence the pareidoliac brings to seeing a face in the clouds. Both are reading presence off of surface. Both are wrong in the same direction for the same reason: the surface doesn’t carry that information, but assholes will hold it up as though it does, which is why dicks like me need to fuck them.

Is there a problem with AI slop? Sure. But the fix for an expectation mismatch is expectation-setting, and Substack already built that feature: the disclosure statement, the “How I make this” note, the writer telling you her process in her own words. That tool addresses the actual problem. The scanner addresses a proxy for the problem, measures the proxy badly in exactly the contested cases, and hands readers a number that claims more than any classifier can know. It flattens intention and creativity into a soulless metric utterly devoid of the context that actually matters, and lets assholes shit all over everything using a Substack-approved badge of shame.

One of these features treats writers as people with processes worth explaining. The other treats a probability distribution as a séance, asking the tokens whether anyone is home.

So scan this post, assholes. Take the number seriously if you want. But understand what you’re holding when you do: not a verdict on whether someone wrote this, only an estimate of whether these sentences sit in a suspicious neighborhood of token-space, produced by an instrument that works best on text nobody disputes and worst on text like mine. There’s a dick here, waiting to fuck the assholes who try it. The scanner has no way to know that.

Neither, from a percentage alone, do you.

Author’s Note: A high level reading of this article might lead one to the supposition that proponents of AI scanning on Substack are represented by Kim Jong Il in this metaphor, or at the very least, by humorless tyrannical authoritarian dictators. This is entirely intentional.