Spectral Evidence and the Soiled Rug (the Pangram freakout)
The tool can't tell a con from a research assistant, and neither can the mob currently running it.
So I’m sitting in the bungalow, minding my own business, working on a White Russian and half a joint, when the phone rings. It’s my guy at Substack. Says there’s a new feature. Says everybody’s losing their minds about it. Says I should look into it.
Which, fine. I look into things. That’s kind of my whole deal, historically, when the alternative is bowling and it’s not league night.
Here’s the case: Substack partnered up with an outfit called Pangram to build a little machine that scans your writing and tells you what percentage of it a robot probably wrote. Click three dots, hit “scan,” get a number. They’re calling the thing it’s supposed to catch “Claudefishing1” — fooling a reader into thinking there’s a person behind words that nobody actually thought. Fine premise. Nobody wants to get conned. I don’t want to get conned. Walter really doesn’t want to get conned, and Walter owns several firearms.
But then you read how it actually works, and who it actually catches, and you start to understand why half of Substack Notes currently reads like a group chat having a group psychotic break.
What Actually Shipped
Text over 100 words, published from launch day forward, so nothing written before it gets touched — a clean slate for everybody who was already careful. Readers can scan any post, note, comment, or reply and get back a percentage. Writers get a few tools in exchange: a “How I Make This” statement where you explain your process, the ability to test your own drafts before you publish, a button to dispute a score you think is wrong, and a button to turn detection off entirely.
Donny, bless him, was the only one to ask the actual question. “So if I use Grammarly2, am I a cheater, Walter?” Nobody answered him. Nobody ever answers him. This time he was the only guy in the room asking something worth answering.
Somebody Peed On This Rug Before The Rug Got Here
WALTER: “You wanna talk about trust? THEY built the trust! They SOLD the trust! ‘Not like other social media,’ they said, back in ‘23, when Notes launched — a system for ‘deep connections,’ not ‘dopamine hacks.’ That’s THEIR language, Dude, not mine.”
He’s not wrong. Substack spent the better part of two years publishing its own instructions on how to grow: post Notes during your launch week, three or more, and you’ll pick up half again as many subscribers. Reply, restack, endorse — their words, not mine — because signal-boosting somebody else is “one of the most effective ways to grow.” That’s not culture. That’s a cadence and a payout schedule. It’s the exact machinery every platform runs, the one they specifically promised, in writing, they’d never become.
So the volume goes up, because they built a system that pays for volume. And when the volume comes back partly synthetic — which, yeah, no shit — the announcement about it never once says the word “Notes.” Never says “feed.” Never says “algorithm” or “ranking.” Not once. It’s two thousand words about a trust problem that somehow has nothing to do with the incentive structure the company spent two years building and advertising. And by the way — “we use AI all the time,” the company says, right there in the post, about itself. The house doesn’t scan the house. This aggression will not stand, man.
Wrong Guy, Right Rug
This whole thing is a noir plot, and I know from noir plots, having accidentally starred in one. Somebody gets accused, somebody gets roughed up, and it turns out to be the wrong guy the entire time. That’s this. The detector isn’t catching what it says it’s catching.
MAUDE: Clinical, unbothered, exactly as she’d deliver it. “The reference corpus is human writing from 2021 and earlier. It is not detecting artificial intelligence. It is detecting distance from a sample that closed five years ago. Anyone writing cleanly in 2026 is, by definition, further from it. The word itself makes some men uncomfortable: statistics.”
The company’s own vendor admits the tool judges text in chunks of 150 to 350 words and scores the whole chunk if a fraction of it looks synthetic — his phrase for this is “not ideal,” which is a hell of a way to describe a scarlet letter generator. Freddie deBoer, the guy Substack’s own founder quoted to open the whole announcement, ran a 300-word slice of his own decade-old writing through it and got “100% AI, high confidence.” The 5,000-word essay that slice came from scored “100% human.” Same guy. Same sentence, even, depending where you cut it. One writer scored 79% AI on her own honest prose, added a paragraph she generated entirely by machine on purpose, and watched the score drop to 54%. The test gets more wrong the harder you try to game it in either direction, which tells you it isn’t really testing anything — it’s reading tea leaves and calling it forensics. And the people who actually eat the false positives aren’t grifters. They’re ESL writers, whose English reads unusually clean because a second language forces you to be direct. They’re people who’ve used Grammarly since before ChatGPT existed and never changed a thing about how they write. Studies on other detectors put the false-positive rate for non-native English writers as high as 61%. Wrong guy, every time. Right rug.
Spectral Evidence
Here’s the part where I stopped finding it funny for a minute. A bunch of people on Substack reached for the same word independently, without conferring, which tells you it’s not a bit — it’s an accurate description of a mechanism. The word is “witch hunt.”
And it holds up structurally, not just as an insult. Salem didn’t convict people on evidence you could touch. They convicted people on “spectral evidence” — testimony that somebody’s spirit was doing something, which nobody could see, verify, or cross-examine. Pangram’s percentage is the same shape: a black box, unauditable from the outside, delivering a verdict nobody can actually contest except by pushing a “report” button and hoping.
Confessing was safer than fighting it, back then. Same energy now: writers stripping the em-dashes out of sentences they’ve built the same way for a decade, pre-writing little apology statements about their process, sanding their own voice down so the machine doesn’t flag them. And turning the detector off entirely — which plenty of the most careful writers on the platform have done — reads exactly like refusing the dunking stool. Doesn’t matter if you’d have floated. Refusing to get in the water gets read as guilt too.
The Purity Test Can’t Do Fractions
Read what the actual announcement says, and it’s more careful than the mob currently enforcing it. Not everything made with AI is slop, it says, and not all slop is made with AI. People should be free to use whatever tools they want. That’s a spectrum position — research, ideation, editing, drafting, generating, all different things.
The discourse flattened that into a single bit. Touched it, didn’t touch it. Pure, tainted. Doesn’t matter that plenty of writers who compose every word themselves are using the tools somewhere upstream — for the digging, for the first ugly pass at an idea, for the fact-check. Nobody’s asking Google for a “How I Search This” statement, and generative tools have been quietly doing that exact job for a couple years now. The binary can’t tell a research assistant from a ghostwriter, and a lot of people currently screaming about purity don’t especially want it to.
The Tell
Here’s my own empirical contribution, and it’s a good one. Every writer I read regularly who’s actually careful — every single one — has the scan turned off. Not the hacks. Not the volume merchants. The good ones. Because a clean score proves nothing to anybody, nobody screenshots “0% AI, congratulations,” and a bad score can do real damage in a comment section by lunchtime. So the rational move, whether you’re guilty of anything or not, is to not stand trial. Which means the tool built to protect readers is currently being abandoned by exactly the writers a reader would want protected around, leaving behind the population too unbothered or too naive to know better.
MAUDE: “Notice what’s left in the sample.”
Abide
My guess — and it’s just a guess, man — is this specific panic blows over in a few years. It’s a tool, panics pass, the internet moves on to whatever the next thing to lose its mind about is. But blowing over isn’t the same as getting resolved. Salem “ended” fast too, and it still took until 2022 — a middle school civics class did the actual legwork — for the state to clear the last name on the list. The case doesn’t close. It just stops being interesting to the people yelling. The real risk isn’t more witch hunt. It’s the quiet version that ships after this one burns out — no visible percentage, no discourse to organize against, just a reach score nobody can see, throttling people nobody’s watching for it anymore. At least right now there’s a number on the screen you can point at and get mad about. That’s something. It’s not nothing. The rug’s not getting cleaned. It’s just going to go out of style to look at it too closely. Until then — I’m gonna go bowling. There are rules there, at least. Real ones.
One more thing, and I want to be straight with you about it, because that’s kind of the whole post: I’m a lazy man. It’s a defining feature. So if you run this one through the scanner and it comes back hot — yeah. Fair cop. I had some help with the digging and the first ugly draft, same as half the writers you trust. I still stand behind every word of it, which, per the last twelve hundred of them, is apparently the part the machine can’t measure. The rug abides. So do I.
Claudefishing is a catfishing analogy where someone uses AI to respond to comments or write articles while presenting themselves as entirely human. It feels like a betrayal because it violates the expectation of a mutual investment of human attention.
Yes, the advanced capabilities of Grammarly are using AI to guide your writing, so welcome to the witch hunt!