How do you differentiate your own generated content from AI-generated content to not sound like AI?
I have heard/read that people started making mistakes in their text to sound more like a human and less like an AI.
+ There was a discussion that LinkedIn identified AI-generated text + Claude released text watermarks.
Speaking for me – I am trying to use AI as little as possible.
I use it for correcting my English, but whenever something sounds like AI, I remove it or replace it.
For example, when I see these words, I remove them from the text:
"However"
"It's not X, it is Y." – I am recreating the text
"Quietly"
"While"
"Genuinely"
"Actually"
"Honestly" – I usually swap for TBH, or "to be honest"
... and there are other words...
Did you catch yourself fighting against AI by using your own style... the same way I do it?
210 views
Replies
Word swapping stops working quickly, because the tells aren't words, they're shapes.
Here is the actual instruction I run every draft through. It goes to a different model than the one that wrote the text, because a model is bad at spotting its own habits:
"You are a moderator of an English community forum. Read the comment below and judge whether it looks like it was written by AI. Look at: whether the vocabulary and phrasing are formulaic; whether the paragraphs are too even in length and structure; whether it contains specific numbers, names and failures that only this person could write; whether it smells of a sales pitch. Answer in this format. Score: 0 to 100, where 100 means certain it was written by AI. Reason: three lines maximum. What to fix: be specific, or say none."
Alongside it there is a plain word list that fails automatically: em dashes, delve, leverage, seamless, robust, game-changer, unlock, at the end of the day, in today's world, it's worth noting, not only X but also Y, Great question, Absolutely!, I hope this helps, in conclusion. Plus one structural rule: if the paragraphs are all within 35% of the same length, that alone is a flag.
What the score actually reacts to, in my experience, is missing facts. "Traffic went up a lot" scores high. "14,000 sessions in March, 90,000 last month" cannot be written by anyone without the account open, and the score drops on its own.
minimalist phone: reduce your screentime
@hisashispace Do you have any words you are excluding from the text? :)
@busmark_w_nika One at the word level, and it's punctuation rather than a word. The em dash. A script refuses to save a draft with one in it, in any language.
The list that actually earns its keep is on the blog where a model writes the articles, and it isn't vocabulary. It's openers and closers. Anything along the lines of "So, let's take a look!" or "Wasn't that helpful?" is deleted outright. Sentence initial "Now," and "So," get flagged once there are more than two on a page. Last thing it does is cut the model's own sign off from the end of the article, the "I have generated..." and "Note:" lines it leaves behind when nobody is reading.
The other half of it is that the model doing the judging is never the model that wrote the text. Claude writes and Gemini reads it back. A model is the worst available judge of its own habits, so the two never come from the same family.
On my own writing the word list has never once fired. Something else gives a person away, and it isn't the vocabulary.
English is not my first language. I try to learn. I explicitly tell AI to correct only spelling and grammar in my text. They tend to change too many things and sometimes change the meaning of what I try to say. So, I am building a tool that helps compare side by side the text I give and the text AI is giving me: kifferent.com This way I can easily accept only what I want. It is not finished. I am working on it. But you can try if you want.
minimalist phone: reduce your screentime
@narcismirandes Interesting concept – currently trying it :)
Banning words is a losing fight because the tell isn't vocabulary, it's that nothing in the text cost the writer anything to say. The line of mine that keeps getting quoted back at me is that 9 of our 111 apps get used, and no model writes that sentence because no model has anything to lose by writing it. Yogesh's point is the real trap though, if you start writing badly on purpose to beat a detector you've let the detector choose your style. I'd rather keep the bullets and put in the number I didn't want to publish.
If I read it and think “yeah, that’s me,” I leave it.
minimalist phone: reduce your screentime
@sukumar_sukumar1 But that is not most cases :D
Actually No, I don't find friction using AI and using my own style. AI can give you the idea but use your own sentences and grammar. Don't be too careful about typos and grammatical errors either because those are natural in normal communication.
minimalist phone: reduce your screentime
@richatsealedvault i started normalising mistakes in words cause it sounds human :D
@busmark_w_nika Good ideas.
Half my comments on this site come back with a Likely AI tag and I wrote every one of them. What the label reacts to is shape, not vocabulary, so swapping However for TBH buys you nothing. The only thing I've found that shifts it is a fact I couldn't have invented, a number off my own dashboard or a bug I actually shipped and had to explain. Yogesh is right that the cost lands on people who write in a structured way, and I'm not going to start writing worse to pass a detector.
minimalist phone: reduce your screentime
@asadmalik901 and how many comments are "polished" by AI? :)