I wrote the AI slop myself

I can’t post anything on LinkedIn because everything I start writing ends up sounding like AI slop. I’ve tried using prompts to strip out the generic phrasing and patterns. The cleanup works, but the result still comes back in a tone that just isn’t me.

I’m writing AI slop by hand -> running it through a prompt -> getting a more human-sounding text -> throwing it in the trash anyway.

My native language writing skills are holding up. The glitch happens at the ‘English + professional post’ combo. Hopefully, this isn’t brain atrophy but just a biased sample that can be improved in a month or two. The model outputs what it was trained on, and my brain training data for professional English posts over the last few years contains too much AI-generated content. And just quitting social media probably isn’t enough here. It cuts off the fresh garbage input, but it won’t retrain the taste on its own.

It used to be that writing in English often felt smarter than writing in my native language — these days it’s the opposite.

Now instead of writing in English from the start, I draft the first version in my native language and then translate it (with AI, of course). But even then, AI loves to sneak slop back in, and it takes a few iterations to land on something that lives between flat Google-Translate literalism and polished AI slop.

You write the whole thing yourself, just want to fix grammar with AI, blink and boom, you’ve got AI slop. The work didn’t go away — instead of fixing grammar, you’re now hunting unwanted edits.

Because if you don’t give it a constrained translation prompt and just say “translate this” or “fix grammar” the model takes the default path. And the default after RLHF is basically “make it good”: smooth it out, make it natural, modern, maybe a little tech-savvy. Which is exactly the high-probability default. Without an explicit “do not improve this,” it improves by default. So the constrained prompt isn’t paranoia. It’s the minimum needed to stop the model from silently editing you while pretending to translate.

More and more people are trying to avoid AI content. Not sounding like AI matters for everyone now, but for non-natives it’s harder. A non-native writer is often afraid of mistakes and asks AI to “fix the English.” As a result, the person loses exactly what they wanted to preserve: their authorial position. So right now I’m testing this pipeline:

1. Write in my native language, keeping it close to how I actually speak.
2. Get a constrained translation — close to the source, no polishing, no added confidence. This is my English baseline.
3. Fix only the parts that sound too literal or unnatural in English, using local edits instead of rewriting the whole thing.
4. Then diff the edited version from step 3 against the baseline from step 2 and flag only where text drifted: more formal, confident, corporate, amplified, optimistic, or smoother than it was.
5. I make the final calls.