Slop Is What Falls Out When Nobody's Steering

Medium sent me an email this week asking what I'd do if I could instantly and flawlessly identify anything written by AI.

Then Zulie, who wrote it, answered her own question. Medium isn't buying an AI detector. She ran the test herself, had Claude write a few paragraphs about her cats, pushed them through Pangram, and got back a verdict of human. What Medium invested in instead is people: a curation team reading stories and handing out distribution labels.

Her argument for that is when readers ask whether something is AI, they aren't really asking that. They're asking whether they can trust it.

She's right. She's also describing a standard that most human writing has never cleared.

Nobody actually wants a detector

A detector answers a question about tooling when the question readers have is about care.

Bad writing predates the transformer by a few thousand years. The listicle assembled from three other listicles. The quarterly report nobody read before sending, the thought leadership post that took four minutes and says nothing. Writing like that comes from an author who doesn't care and a publish button that doesn't ask. No model required.

Run detection at 100% accuracy tomorrow and all of it still exists. You've filtered on the wrong variable.

The variable that matters is whether anyone decided anything. Did a person live this, take a position, and put their name on the result? That question survives every change in tooling, which is more than detection can say.

I haven't written unassisted in years

I don't write well. I never have. I think out loud, in fragments, the way I talk, and that has never translated cleanly to a page. For years Grammarly sat between my brain and everything I published. It showed me alternatives, I picked one, and that was the sentence.

Nobody ever called that slop. Nobody asked me to disclose it. Grammarly shipped in Word and Gmail and half the internet's writing ran through it, and whether the resulting thought was still mine never came up.

It never came up because the answer was obvious. Grammarly never had a position and never decided what was worth saying. It looked at meaning I'd already committed to and offered a cleaner container for it. Every accept was a judgment. Every reject was a judgment. That's editing, and editing has been part of authorship for as long as authorship has existed.

Now I dictate. I talk at an agent the way I'd talk at a colleague, it catches the thought, and we shape it together. Claude is my editor. BlackOps is my manager, the thing that holds the voice spec and refuses drafts that drift off it. This post is that. It started as me being annoyed at an email, out loud, in a conversation, and it got molded into what you're reading after a lot of back and forth.

That's faster than I used to work and it's still my thought. I'd bet money it's already how a large share of what you read this week got made.

You can't see how much correcting happened

You can usually tell when a model wrote something. There's an accent. The same sentence shape inverted over and over, the colon that promises a reveal, three parallel fragments in a row, a punchline at the end of every paragraph. It's recognizable, and anybody who reads a lot has learned to spot it.

Zulie would push back here. She says the individual tells rotate, that every one of them has its moment and gets replaced. She's right about the surface. The specific words age out fast. What doesn't age out is the shape underneath them, the sense that the sentences are arriving in a rhythm nobody chose.

So the accent is real. What you can't see from the finished page is how much correcting happened before it got there. Whether I wrote every word by hand at 2am, or ran a pipeline and killed four drafts, or sat there striking out lines one at a time until it said what I meant.

That's the part that decides whether it's any good, and it's the part the page hides.

Which means catching the accent doesn't tell you what people think it tells them. It tells you nobody was minding the output. Those look the same on the page and only one of them is the problem.

So the honest framing isn't human versus machine. It's owned versus unowned. Did someone stand behind this, or did it just get emitted?

A human can emit. I've watched people ship unowned writing their whole careers. An agent running under enforced rules can produce something a person will defend in public with their name on it.

My doctor talked to his phone

I had a checkup recently. My doctor talked to his phone through the whole visit. At the end I got a summary of my health, written by a model, from a recording of a conversation.

He didn't write a word of it. It is still entirely his assessment of my body, his diagnosis, his instructions, and his license on the line if any of it is wrong.

Is that slop? Obviously not. It's the most owned document I received that month.

What makes it trustworthy is the accountability, not the keystrokes. Somebody read it, somebody stands behind it, somebody pays if it's wrong. That test works for medical records and it works for writing, and it worked before any of this existed.

Twenty rules and I'm still correcting it

I'll show you mine, because assertions about care are cheap and everyone makes them.

Everything I publish starts as a seed rather than a prompt: something that came out of real work, logged the moment it happened, with the context still attached. This post started as an email that annoyed me and a position I already held.

From there it runs a chain. Draft, internal linking, SEO, media, then a proofread against a voice spec that exists as an actual file, then a variant pass for wherever it's landing. Every stage is a gate, and gates fail. Drafts die at the proofreader more often than they pass on the first run.

The spec is around twenty rules. No hedging, no em dashes, don't run three parallel sentences in a row, section headings have to be plain and lifted from a line in the body. On top of it sits an eval gate that grades the output against a separate rubric, because the thing that wrote the draft cannot be trusted to grade its own work.

This post went through all of it and passed clean.

Then I read it again and found a sentence claiming slop is "engineered" to satisfy readers. Nothing engineers slop. It's what falls out when nobody's steering. The sentence sounded like an argument and wasn't one, and it cleared twenty rules and an eval gate without anything catching it.

I caught it because I was reading. That's the only reason.

So no, the machines are not good at this. That's why slop exists. They produce the accent by default and they'll produce it forever, and the rules catch a lot of it and never all of it. The last pass is a person reading every line and cutting the ones that don't hold. There's no arguing about it. It doesn't get a vote.

If I wouldn't defend a sentence in a room full of people who disagree with me, it doesn't go out. That is more deliberate than what most writers do, which is type into an editor at 11pm and hit publish while the thought is still warm.

Additional hands

Additional hands. Additional eyes. Additional thinking. All running on your rules.

The rules are the whole product. Take them away and you have a text generator, and text generators produce slop by default because default is what they're for. Put them back and you get the same position you already held, expressed more times, in more places, with fewer errors than you'd catch alone at midnight.

Curation can't tell you that

Medium would rather cultivate writing worth reading than play detection whack-a-mole. That's the right call, and they got to it honestly. Zulie's last section says writing started suffering long before AI, then walks through SEO keyword stuffing and viral clickbait as earlier rounds of the same disease: writing aimed at a machine instead of a person. I came in ready to tell them the problem predates ChatGPT and found they had written it down first.

Where I'd push is on what curation can actually deliver. Sorting feeds tells you a human recommended this piece. It doesn't tell you a human minded it, line by line, before it shipped. Those are different questions and only the writer can answer the second one.

She also tells writers to stop being afraid of em dashes, and she's right that writing to beat a detector is its own kind of rot. My spec bans them anyway, not to pass as human but because I got tired of reading my own sentences in a cadence I never chose. Cutting them is faster than arguing about which ones earned their place.

That's the point. The rules are mine. If they were somebody else's scanner, she'd be right to tell me to drop them.

The burden lands on the writer, which is where it belongs. Write something you'd sign. Enforce your rules on every word that ships under your name, whoever or whatever typed it, and kill the drafts that don't clear the bar.

Do that and detection is irrelevant, because the reader gets what they were actually asking for. They were never asking who typed it. They were asking if anyone was home.