Last week I asked an AI assistant to write our launch post. It gave me three paragraphs in about forty seconds. Two made it out of the drafts folder. My cofounder read the final version and asked, half joking, if we had quietly started using an AI writing tool in 2024. We hadn't.
I did exactly that. I pasted the output in and changed a few adjectives. The piece shipped. It read like it was written by someone who had never talked to a customer.
The fix starts with the first sentence, which is also the most underrated line in any draft. AI search engines and excerpts pull the first direct answer, and AI-written openers tend to be broad claims, so the answer gets quoted before you get a chance to say anything true.
How to fix it
Score the draft before you publish. I use a 20-tell scorecard, grouped by structure, language, and evidence, with a 0 to 2 scale per tell. Anything you cannot verify scores 0.
Then run a cleanup pass:
- Audit. Note every tell scoring 2.
- Cut the weak sections. AI almost always writes one more than the piece needs.
- Add one specific detail the model could not know. A number, a name, a date, a real quote.
The rewrite
You do not need to regenerate the whole piece. Swap the flagged lines using plain language:
"Content is king in the AI era" becomes "Start with your first real point."
"Our platform transforms the way teams work" becomes "It classifies support tickets and drafts replies. Setup takes an afternoon."
The tells
The scorecard sorts tells into structure, language, and evidence. Structure catches the rhythm problems. Language catches the vocabulary. Evidence catches the hand-waving.
If you want the full scorecard, the rewrite table, and the voice sheet, I put together The AI-Slop Audit: a 20-tell audit you can run in about 15 minutes. Check it out at https://probiex.gumroad.com/l/ai-slop-audit.
How do I know if my AI-written draft is bad?
Score it against the 20 tells; a result from 17 to 34 means it reads as AI and needs a rewrite before you publish.
How do I make AI writing sound less generic?
Add the specifics the model could not know, then run the flagged phrases through the rewrite table.
Last updated October 2026 · Probiex
