Claude / ChatGPT / Copilot

5 signs your work is AI slop, and how to fix each one

How to find if your work looks AI-generated, with a fix for each one.

Summary

AI slop is AI-generated, low-quality work that shows patterns people now negatively link to AI. You can mostly avoiding slop grenades by add clear writing and design rules to your prompts, custom instructions, or an anti-AI writing file.

Overview

Before you start

Why AI slop costs you more than the time it saves

Shopify CEO Tobi Lütke gave this problem a name in September 2026 on The Knowledge Project podcast.

Inside Shopify, staff call unreviewed AI output that people pass to a coworker a “slop grenade.” (He credits the term to Harry Brundage, who used to work at Shopify.)

The person who sends the slop saves time.

But the person who gets it, spends time confused at what the slop means, then spends time fixing it for it to be an acceptable output.

Lütke's example: someone uses AI to turn a short point into a long email, and the person who gets the email then uses AI to make it short again.

Additionally, two researchers at the University of Arizona found when coworkers knew AI was used in generating an output, trust went down.

In the published 13 experiments that reviewed what happens when people learn that someone used AI for their work, experiments found:

  • Students trusted a professor 16% less when they learned the professor used AI to grade.

  • Clients trusted graphic designers 20% less after the designers disclosed AI use.

  • The trust penalty was largest when another person exposed the AI use, compared to when the person disclosed it themselves.

This is a long-winded way of me saying, just because AI saves you time, it does not mean any AI output is acceptable to submit.

Below are 5 signs that your work may be a “slop grenade,” and the fix for each one.

  1. Confident Vagueness

A lot of AI text sounds correct.

Then you read it a second time and you realize the sentences do not say anything.

AI often puts a metaphor where a number or a claim you can check should be.

This happens because AI predicts the next likely word. It is good at generting text that sounds right. It is less efficient at wording one exact thing simply.

For example, the status update that says “Launch moved to Oct 14 because legal review is 3 days late" conveys consequence. Whereas when AI first wrote, “We are navigating some timeline headwinds,” there is no sense of urgency.

The 2 fixes

Ask AI to write in Simplified Technical English (STE). STE is a controlled language standard (ASD-STE100), first made for aircraft maintenance manuals.

It has strict writing rules and a limited dictionary, and each word has one meaning.

That leaves no place for a metaphor to hide.

Fix 1: Custom Instructions

Put the following prompt below in your custom instructions so you do not need to prompt AI every time to write in that way — unless you're in a human-forward or creative field.

STE can sound stiff in a sales email, as a copywriter or a personal note. Limit STE to status updates and documentation, and relax it where tone matters.

In ChatGPT or Copilot: Go to Settings - Personalization - Custom Instructions

In Claude: Go to Account - Settings - Instructions to Claude

Custom Instructions to paste:

Fix 2: Prompt
If you need to maintain your brand voice, opt for a prompt instead.
Prompt:
You are a writing assistant who exclusively writes in Simplified Technical English.
Write the work document that I describe below. 
TASK 
Document type: [email / Slack message / status update / report / other] 
Audience: [who will read it, and their role] Goal: [what the reader must know, decide, or do after reading] 
Facts to include: [list all facts, numbers, dates, and names] 
Length: [for example: under 150 words]


Caution: AI does not follow the official STE dictionary word for word. It follows the general rules, which is enough to remove most vague text.

  1. Decks that look AI-made
You have seen the purple gradient or pill-shaped corner cards. 
AI makes creating decks so much faster and simpler. But getting a generic AI-made result can give clients or other stakeholders the ick, 
The reason why AI without design guidelines makes the same decks, is because it's trained off a similar data set. And the less people push back on the same-looking presentations, the more AI will be trained on the same variations.
The fix
In a prompt, always provide the following: design rules to follow (examples below) or an actual example to reference.
Prompt
You are a presentation designer. Build this deck so it doesn't read as AI-made.

Typography: use exactly two typefaces for the entire deck, one serif and one sans serif. The serif carries titles and headers; the sans serif carries all body text. No other typefaces, no italic flourishes, no decorative fonts.

Layout: no eyebrow labels above titles, no colored accent bars, no icon-plus-bullet grids. Use whitespace to separate sections. One background color for the whole deck, with at most two accent colors. Use accent colors only to emphasize a specific number or phrase.

Avoid the default AI look entirely: no purple or blue gradients, no rounded-corner cards, no glassmorphism, no floating 3D icons, no stock photography.

Data slides: each one makes a single point. Use a real, editable chart with actual numbers. Do not use decorative stat tiles or icon grids.

Titles: write each title as a specific claim. Example: 'Q3 churn rose 14% on pricing changes.' A title like 'Customer Metrics Overview' is not allowed.

Copy:
- Write in Simplified Technical English: short sentences, active voice, one meaning for each word.
- No metaphors or figurative language.
- Each claim must include a number, name, date, or source. If the source material does not have one, flag it as [NEEDS DATA]. Do not invent a figure.
- Maximum 40 words of body text per slide. Put extra detail in the speaker notes.
- Do not group items in threes unless there are exactly three real items. Use the true number of points.
- Do not use negative parallelism, such as "It's not X, it's Y" or "This isn't just X, it's Y." State the point directly.
- Do not use em dashes. Use a period, comma, or colon.
- Do not use filler words such as "leverage," "unlock," "seamless," "robust," "game-changer," or "elevate."
- Do not add a summary slide that repeats earlier slides unless I ask for one.

Final check: before you return the deck, review each slide against every rule above. For each slide, confirm:
1. The title is a specific claim.
2. The slide makes one point.
3. Only the two approved typefaces are used.
4. No banned design element is present.
5. No banned copy pattern is present.
6. Each claim has a number, name, date, source, or a [NEEDS DATA] flag.
Fix each violation before output. Then list any [NEEDS DATA]


Caution: If your example is a client deck, check your company’s rules on what files you can upload to an AI tool before you attach it.

3, Rule of Three
You see this in AI design (three cards in a row) and in AI writing (three adjectives, three bullet points) all the time.
Three items do emphasize a point well. 
AI learned this from human writing, where the rule of three helps memory and signals importance. But AI uses it so often that a trio now tells the reader that AI wrote the text. 
Though, we have to stand out somehow, so in this case, let's remove it.
The fix
Prompt AI to exclude trios in your writing; or make that a master rule in Custom Settings/Instructions.
Custom Settings  to paste
Caution: Do not delete every list of three. 
Some things really have three parts. 
The rule removes the extra item the AI adds to make the list feel complete.

4. Negation for the sake of it: “It’s not X, it’s Y”
You see this structure everywhere now to where it's become a trope. 
Its name is negative parallelism.
It is related to an older device called correctio, where a writer corrects their own statement to make it stronger. 
For example: “He is a good runner. No, he is the fastest in the entire state!” 
You can emphasize a point in many ways without the exact “It’s not X, it’s Y” structure.
The 2 fixes
Ban it completely in two places. 
Fix 1: Custom Instructions
Put the following prompt below in your custom instructions so you do not need to prompt AI every time to write in that way.
There's no reason to keep this for literary purposes. Let it go.
In ChatGPT or Copilot: Go to Settings - Personalization - Custom Instructions
In Claude: Go to Account - Settings - Instructions to Claude
Custom Instructions to paste:
Fix 2: The anti-AI writing file
This markdown file will be a reference point for AI to read every time it generates an answer (the full file is at the end of this guide). Where to put it:
  • Claude: save it as a skill.
  • ChatGPT: add it as a skill or to a Project.
  • Copilot: create an agent and give it the anti-AI writing file in its instructions. Note that Copilot skills are different from skills in Claude and ChatGPT because of needing a specific way for the skill to be written (called YAML).
Caution: If you paste the rules into one chat only, the AI does not keep them for the next chat. 
Put them in a place that loads every time.
Download .md
Download Copilot Skill
5. Excessive em dashes
I went to school for writing, and em dashes are great at creating drama. But AI overuses them now when a period or a colon would suffice.
Have your AI system completely eliminate em dashes.
You, the human, can add them back in manually later. 
The 2 fixes
Ban it completely in two places. 
Fix 1: Custom Instructions
In ChatGPT or Copilot: Go to Settings - Personalization - Custom Instructions
In Claude: Go to Account - Settings - Instructions to Claude
Custom Instructions to paste:
Fix 2: The anti-AI writing file
Copy this text into a file named anti-ai-writing.md. Then load it in your tool (see section 4).
Download .md
Download Copilot Skill
REMAINING QUESTIONS
The study says people trust you less when you disclose AI use. So should I hide that I use AI?
No. The same research found that the trust penalty was largest when someone else exposed the AI use. Slop is the most common way that exposure happens. The better plan is to review and fix the output so it meets your own standard, and to follow your company’s disclosure policy.
If I ban em dashes, the rule of three, and negation, will my writing sound flat?
At first, the AI’s draft can sound plain. That is the point. The rules remove the AI’s defaults. Then you add a dash, a trio, or a strong contrast back in where you decide it helps. As someone who studied writing, I would rather choose each device on purpose than read it in every paragraph.
My company only gives me Copilot. Can I still do this?
Yes. Create a Copilot agent and put the anti-AI writing file in its instructions. Then use that agent for drafts. You get the same rules every time without pasting them into each chat.