I Built an AI Slop Remover

How to Humanise AI Writing: A Free Tool and Term List

Downloadable Claude Skill and Custom GPT

On 21 July, Substack switched on an AI detector. It’s built with a company called Pangram, and it scans posts, notes and comments over 100 words. Anyone can run it. Readers on your work, you on your own drafts, or a stranger on someone they’ve taken against.

It’s opt-in and there’s no penalty attached. Nobody gets ranked down or banned for a high score. Chris Best has been careful to frame it as a first step toward transparency rather than a rule about which tools you’re allowed to use.

I’m relaxed about it, and I should explain why before I get to the interesting part.

I’ve never hidden that I use AI. I wrote an Honest AI Manifesto in April saying exactly how and where. Some pieces here are drafted almost entirely by hand and tidied with a model. Some start as a conversation and get rewritten by me four times. A few are mostly mine with a paragraph I couldn’t crack handed over in frustration. That’s the honest answer and it’s been on the record for months.

So a scanner isn’t a threat to me. It might tell readers something they already know.

I do have one worry, and it isn’t about my score. The false positives land in a pattern. They hit people writing in a second language, and they hit neurodiverse writers whose rhythm and repetition a model reads as synthetic. One writer ran the same piece twice last week and got 100% AI, then 100% human. That’s a coin flipper, not a detector.

But that’s an argument for treating the number with suspicion, not for hiding from it.

Slop is a readability problem before it’s a moral one

Using AI is fine. Publishing what it hands you first time is where it goes wrong.

Generated prose has a texture. Every sentence lands between 15 and 20 words. Every section closes with a tidy summary. The same eighty verbs turn up over and over. Delve. Leverage. Foster. Navigate. Nothing is technically wrong, and that’s the trouble, because there’s nothing to catch on.

Human writing is lumpy. Two-word sentences next to forty-word ones. A paragraph that goes on too long because the writer got interested. A joke that doesn’t quite land. Repetition where they should have varied it. Those are the warts, and readers need them. They’re the handholds.

Strip them out and you get text that’s correct and frictionless, which sounds like a compliment until you notice nobody finished reading it.

So the case for humanising your drafts has nothing to do with detectors. Slop is boring, and boring loses subscribers faster than any label ever will.

What I built

A tool that scans a draft, tells you what’s wrong with it, then edits. You paste your writing in. That’s the whole interface.

The scan counts rather than guesses. Ask any AI to tell you if your writing sounds robotic and it’ll have an opinion, and the opinion will be different tomorrow. This one comes back with numbers, and they’re the same numbers every time.

It catches what your eye slides past. Em dashes. Semicolons. Roughly 300 words and phrases that show up constantly in generated text and hardly ever in good writing. It gets the structural habits too, like the self-posed question (”The result? Devastating.”) and the trailing clause that adds nothing (”...solidifying its position as a market leader”).

Then it measures rhythm. It works out how much your sentence lengths vary, and your paragraph lengths, and tells you when the answer is “not enough”.

You get a number at the end. Under 10 per thousand words reads human. Above 30 and the piece needs rebuilding rather than tidying.

I ran a deliberately awful sample through it and got 48. I ran a passage I was proud of and got zero.

What it won’t do

It won’t help you beat Pangram, and I’ve built it to refuse if you ask.

The refusal is practical. The job can’t be done reliably. Detectors score statistical patterns in how text was generated, which isn’t the same thing as the vocabulary and sentence shapes this edits. The two overlap a bit. Your score might drop after an edit. It might go up. Anyone selling you a guaranteed number is selling you a guess in a confident voice.

And chasing the score is the wrong game anyway. If you’re editing to satisfy a classifier, you’ll produce writing that satisfies a classifier. Nobody has ever subscribed to a newsletter because it scored well on anything.

Edit so it reads better. The score does what it does.

The part many ‘humanising’ tools get wrong

Every de-slop tool I’ve tried is too keen.

You write about actual landscapes and it strips out “landscape”. You use “systemic” correctly in a piece about systemic risk and it swaps in something vaguer. You make a joke and it sands the joke off, because jokes are irregular and irregular looks like a mistake.

What comes back is clean, correct and dead. Which is the same problem you started with, wearing a different coat.

So this one shows its work. Every edit comes with a change log, and underneath that, a list of everything it flagged and deliberately kept, with reasons. If it left your weird metaphor alone, it says so and says why. You can disagree with it, which is more than most editing tools allow.

Where to get it

It’s a Claude Skill, so it loads by itself when you mention humanising a draft. There’s a ChatGPT version built as a custom GPT. It runs the same scanner under Code Interpreter. The Gemini version works but loses the diagnostic, because Gems can’t run code. I’d rather say that now than have you find out.

The term list is a plain text file. Add your own words, delete mine, and the scanner picks up the change without anyone touching code. My list bans “delve” and “tapestry” and about 298 other things. Yours will be different. Mine bans “utilise” purely out of spite.

Download the files for Claude and ChatGPT, including instructions, here (zip file - unzip when downloaded) …