How to make AI writing sound like you
You don't need to start over. Most chatbot drafts only need five or six lines changed. Here is how to find them and what to put in their place.
If you used ChatGPT for an email and didn't like how it came out, you don't have to start over. Keep the draft. Cut the first and last sentence, put your own facts where it's vague, and read it out loud to see which lines you'd never say. Most of the draft is fine. The problem is usually five or six lines, and those are the ones to change.
Start with the first and last sentence
The first one tends to set a scene about the modern world and the last one tends to tell the reader what the email meant. Take both out and look at what's left. Nine times out of ten nothing is missing, and the email now starts on the actual point, which is where you would have started it yourself.
Then put in what only you know
Next, find the most vague claim in the draft and replace it with a fact. The model can't know your client's name or your numbers, so it writes around them, and that is a lot of why it sounds like everybody else's email. If sign-ups went from 190 to 312, say that instead of calling it a significant increase. If the draft praises your leadership skills, write down what you led and how many people were on it.
Read it out loud
Read the draft as if the person were in the room with you. Some lines are hard to say to someone's face, like telling them you wanted to take a moment to reach out. You'd just say you have a quick question about the proposal, and that they can call you after 2 if anything looks off. We think this catches more than a word list does, since a word list only checks words and this checks whether it sounds like you.
Don't worry much about single words
There's a lot of worry online about particular words. People see one fancy word in a post and decide a chatbot wrote it, which isn't fair to whoever wrote it. People used those words long before chatbots did.
Our own first version made the same mistake. It marked single stock words, and on a test of 3,000 texts that people wrote, it showed stock phrases on 24.7% of them. So we changed it. Now a rule only counts when three or more different habits show up in the same stretch of about 300 words, and on those same texts it dropped to 1.8%.
What makes a passage read as machine-made is several habits packed into a small space. A big opening line, a neat list, a long dash, a contrast that argues with a point nobody made, and a lesson at the end. One of these on its own is just writing. Four or five in one paragraph is a pattern, and readers pick up on it even when they couldn't say what it is. So when you edit, look at a whole paragraph at a time and break up whatever is piling up there.
Asking the chatbot to fix it
The obvious move is to paste the draft back into the chat and ask it to sound less like AI. In our tries this goes badly. The model rewrites the whole thing, so your good lines go out along with the bad ones and a new set of stock phrases comes in. You end up with a different draft that has the same problem.
It works better if you tell the model which lines are the problem and why, and ask it to keep every fact and leave the rest of the text alone.
Where Desloppy comes in
Finding those lines in your own draft is the hard part, and it's the step we built Desloppy for. You paste a draft, or upload a Word, PowerPoint or PDF file, and it marks the lines that sound like a chatbot, gives a short reason for each, and suggests a fix. You also get a Slop score, and after you edit you can check again and see the two numbers side by side. On drafts under 80 words, fixing every marked line brings the score down about 22 points on average in our tests. Your text is never stored.
Desloppy doesn't rewrite anything for you, and that's on purpose. You make the edits yourself, or you give only the flagged lines to your own assistant along with the real details.
The Slop score is a probability and not proof. It describes how the text sounds and never who wrote it, and short or very plain writing gets misread more often. It also isn't a way to get past someone's checker. The point is to send things that sound like you wrote them, since those are the ones people answer.
You can try 1,000 words free, with no account.