There are two free tools that can strip the robot out of your writing in about a minute. Almost nobody I know uses them. So I open LinkedIn every morning and scroll through the same gray paste, post after post, all of it sounding like it came off the same conveyor belt. The conveyor belt is mostly ChatGPT, sometimes Claude, and you forgot to turn off the part that makes you sound like a press release.
I am talking about the Humanizer skill from blader and stop-slop from Hardik Pandya. Both are open source and both catalog the exact patterns that make AI writing detectable. Both are sitting on GitHub right now with thousands of stars and approximately zero of your coworkers running them. I'm going to rant for a while, then give you the patterns, so you have no excuse left.
Your slop has a fingerprint
A language model predicts the next likely word. The output drifts toward the most statistically average phrasing that fits the most situations. Average is the enemy of voice. That is why every AI post smells the same even when the topic is different.
The good news is that average leaves fingerprints. Once you learn them you cannot unsee them, and you start catching them in your own drafts, which is the entire point.
The tells, in order of how much they give you away
The single biggest one is the fake contrast. It's not just X, it's Y. It's not about the code, it's about the craft. This isn't a feature, it's a philosophy. A real person uses that structure once in a blue moon. A model reaches for it constantly because it feels insightful while saying almost nothing. Every time you write it, state Y and delete the rest.
Next, significance inflation. The moment something ordinary "stands as a testament to" or "marks a pivotal moment in the evolving landscape," a robot is puffing up a normal fact into a TED talk. A 2019 product launch is a 2019 product launch. It does not underscore the enduring legacy of anything.
Then the rule of three. AI cannot resist groups of three because three sounds complete. Innovation, inspiration, and insight. Faster, smarter, and more scalable. Count the triples in the next post you read. Real writing uses two items, or four, or one, because real thoughts do not arrive pre-packaged in threes.
The em dash deserves its own paragraph. I do not use them, partly on principle and partly because they have become the clearest watermark of generated text. When a sentence has three of them stacked up like a person who cannot commit to a full stop, a model wrote it. Use a comma. Use a period. Pick a side.
Signposting is the tutorial-script tell. "Let's dive in." "Here's what you need to know." "Let's break this down." The model announces the work instead of doing the work. Cut the announcement and start with the actual sentence. Nobody needs a trailer before a 200-word post.
A few more to keep in your back pocket. Copula avoidance, where nothing is allowed to simply "be" something. The gallery does not "feature" four rooms, it has four rooms. Vague attribution, where "experts believe" and "industry observers note" do the work of an actual source. The generic uplift ending, where the post closes on "exciting times ahead" instead of a real point. And sycophancy, the "great question, you're absolutely right" reflex that leaks straight out of the chat window into the comment you just pasted.
And lastly, the word "silently" is creeping into a lot of Claude writing. The clankers are starting to steal our good words.
It's just two minutes of work
The fix is not hard and it makes me want to throw away my laptop when I see peoples posts. Humanizer ships 33 of these patterns with a before and after for every one. Stop-slop hands you a checklist and a scoring rubric. You point your editor at one file, paste your draft, and the slop comes out the other side sounding like a person. One minute of setup and a few seconds per draft.
People will spend forty minutes prompting a model to write a 300-word post and zero seconds editing the result. They treat the first draft as the final draft because it looks finished. Finished and good are not the same thing. The model gave you the most average possible version of your idea. Your job was to put yourself back in. You skipped it, and now your personal brand sounds like a slop rocket.
Loosing our writing style
None of these patterns are bad. They are good tools humans have used forever. The em dash is a beautiful piece of punctuation. A short, punchy sentence builds emphasis. A well-placed trio has rhythm. Used sparingly by someone who means it, every one of these moves works.
The model just overuses them. It takes the devices a good writer reaches for once a page and slams them onto every line until they brake.
Just spend the time
Using AI to write is fine. I do it every day. Shipping the raw output is a crime, because it is designed by math to be the least surprising thing anyone could say. The least surprising is a terrible thing to be on a feed full of people trying to sound least surprising.
The tools to fix it are free, they are good, and they are right there. Run them, or keep posting in a voice that belongs to nobody.
Frequently asked questions
How do I know if my own writing sounds like AI?
Read it out loud. If every sentence is the same mid-length cadence, if you keep hitting groups of three, if you find em dashes everywhere, and if a paragraph opens with Let's dive in or In today's landscape, a model wrote most of it. The strongest single tell is the It's not just X, it's Y construction. Humans use it once a year. Models use it once a paragraph. Look for clusters, not single hits. One em dash means nothing. Em dashes plus rule-of-three plus a vibrant tapestry is a confession.
What are the free skills mentioned in the article?
Two open source skill files for Claude and similar tools. Humanizer by blader catalogs 33 patterns pulled from Wikipedia's Signs of AI writing guide, with before and after examples for each. Stop-slop by Hardik Pandya is a tighter set of rules that kills filler phrases, breaks formulaic structures, and forces active voice. Both are free, both are MIT licensed, and both take about a minute to point your editor at. The point of the article is that they exist and almost nobody bothers to run them.
Is it wrong to use AI to write at all?
No. Using AI to draft, restructure, or unblock yourself is fine and I do it constantly. The problem is shipping the raw output without editing it. A model gives you the most statistically average phrasing for the widest set of cases, which is the opposite of a voice. The fix is not to stop using AI. The fix is to treat the first draft as a draft, strip the tells, and put a human point of view back in before you hit post.
Why does AI default to these specific patterns?
Language models predict the next most likely token. The result drifts toward the most statistically probable phrasing that fits the widest variety of cases, which is exactly how you end up with significance inflation, hedging, and tidy three-item lists. None of it is wrong. All of it is average. Average reads as fake because real people have uneven rhythm, strong opinions, and weird specific details that a probability machine smooths away.
