AI slop: what it is, how to spot it fast, and how to fix it

What AI slop is

Slop has a texture you can learn to feel. It is over-polished: every sentence is grammatically clean and none of them say anything a competitor's page could not say. It is evenly paced, with each sentence running to about the same length, so the whole thing hums at one frequency. It is generic, because the model wrote the average of everything it has read on the topic. And it is hedged, so nothing in it can be wrong and nothing in it can be useful.

A draft you asked a model for, then cut, corrected and filled with your own specifics, reads as yours. Slop is the draft that skipped that second half. The model did what it does, the person pasted it, and the reader was left to do the editing in their head. They decline, and they leave.

I write with models every day. The editing pass after the model stops is where the page gets made.

AI writing tells, by model

Each model has a signature. Once you know the signature, you can usually name the model from one paragraph.

Model	The fingerprint
ChatGPT	Reaches for "delve" and stacks long dashes where a comma or a period belongs
Claude	Opens with "Great question" and an exclamation mark, then answers
Gemini	Lists everything in threes: three benefits, three steps, three examples, whether or not the content has three of anything

Beyond the model signatures, five patterns show up whichever model wrote the draft:

1. Lists of exactly three, by reflex. Two options become three. Six steps get squeezed into three.
2. The sentence that denies one thing to assert another. It sounds like insight and contains none, because the denied thing was never on the table.
3. An opening line that announces a lesson is on its way before any lesson has been given.
4. Hedges on every claim: "can help", "may", "in many cases", "it is worth noting". The writer is protecting the sentence instead of informing you.
5. A closing paragraph that repeats the article, which nobody asked for.

Add even pacing to that list and you have the whole kit. Real writing has a four-word sentence next to a twenty-five-word one. Slop has fourteen words, then sixteen, then fifteen.

The 90-second slop test

Paste the draft into a plain text editor and count two things.

First, the banned words. Scan for the drama verbs, the corporate nouns, the inflated adjectives and the stiff academic transitions. Every hit is a point. The full list is chapter 3 of the guide, and it keeps growing as the words change.

Second, the patterns. Lists of three, the deny-and-assert sentence, the announced lesson, the hedge, the summary paragraph, the long dashes, a sentence length that never varies. Every hit is a point.

Add the points up. The scoring bands are in chapter 2, but you can guess the shape: a clean draft scores near zero, and anything in double digits will be identified as generated by the second sentence, whatever the byline says. The whole thing takes about ninety seconds once you have done it a few times.

Why AI slop costs money

Trust goes first. A reader who decides a page was generated also decides nobody checked it, and from then on every claim on it is discounted. They skim, and they leave without doing the thing the page was for.

Replies go next. A cold email that reads generated gets archived, because the recipient can tell you spent nothing on it. The same applies to a LinkedIn post, a support reply, a proposal.

Conversions go last, and quietly. A landing page full of slop still loads, still ranks for something, still gets visits. It just gets fewer sign-ups than the same page written by a person with an opinion, and you never see the visitors who left.

How to humanize AI text: the four-move replacement strategy

The fix is a replacement strategy with four moves, applied in order.

Cut first. Delete the opener that announces the point, the summary at the end, the hedge on every sentence, and any list of three where the content has two or five.

Replace the words. Every drama verb becomes the plain verb. Every inflated adjective is removed, or replaced with a fact that makes it unnecessary.

Add a specific. A name, a number you can stand behind, a tool, a date, a thing that happened. One specific per paragraph is enough to make it yours.

Vary the length. Read it aloud. Where three sentences in a row land at the same length, cut one down to four words or join two together.

Applied to a piece of landing-page copy for a made-up scheduling app:

Before

After

In an increasingly competitive market, businesses of all sizes need to work smarter, faster and more efficiently than ever before. This innovative platform is designed to help teams streamline their workflows, boost productivity and drive meaningful results. Whether you are a small startup or an established enterprise, the solution adapts to your unique needs. Ultimately, success comes down to having the right tools in place.

Rota builds next week's shift schedule from this week's, then flags the gaps you still have to fill by hand. Setup takes one afternoon. It runs on a phone, because that is where shift managers are when the schedule breaks.

The before paragraph could sit on any software site. The after paragraph could only sit on this one. That is the whole test.

How to stop slop at the source with custom instructions

Custom instructions are a settings field the model reads at the start of every conversation, before your first message. Whatever you write there applies to every chat without being pasted again. It is where the ban list belongs.

Paste the words you never want, the patterns you never want, and a line about the sentence length you do want. In ChatGPT the field lives in settings, and my walkthrough of ChatGPT custom instructions shows exactly where it is and what to write. Claude and Gemini each have an equivalent field.

Each model needs its own version, because each has its own default habits to override. The guide ships one for each. Paste it once and the drafts arrive cleaner, which means the 90-second test finds less and the editing pass gets shorter.

The rules then sit underneath every prompt you run afterwards, including anything you take from my library pages for ChatGPT prompts, Claude prompts and Gemini prompts.

The 10-point pre-publish audit

The audit is a checklist you run on the final draft, after the edits and before the publish button. Ten points, each a yes or a no. Four of them:

• Does the first paragraph answer the question in the title, in plain sentences, without a warm-up?
• Is there one specific in every section that a competitor's page could not have?
• Does any sentence deny one thing to assert another?
• Could the last paragraph be deleted without losing anything?

The rest cover the words, the rhythm, the hedges and the byline. Two minutes, every page. If you want to see the standard I hold my own prompt library to, the documentation page explains how the library is built and scored.