“But I wrote this myself!”
You wrote a scene. Polished it for nights, every word is yours. Then, just to be safe, you run it through an AI detector — and the thing claims: “100% AI-generated.” It stings. Because what if a publisher, a jury, or a professor runs the same test and doesn’t believe you?
We hear this fear more and more often. So we sat down and took the problem apart properly — with real tests against a real detector. What came out of it surprised us. An honest workshop report.
What an AI detector actually measures
First we had to understand what such a detector is even looking for. We worked with Pangram, one of the more accurate ones on the market.
Older detectors count “perplexity” — roughly: how surprising the word choices are. The problem: by that logic the Bible, the Declaration of Independence and “Frankenstein” all count as AI, because language models have read these texts thousands of times and find every word “expected.” Useless.
Pangram does it differently. It is a neural network trained on “synthetic mirrors”: for every human text, an AI text with the same content and tone was generated, and the model learned to hear out the machine’s fingerprint — not a single feature, but the overall statistical pattern of AI language.
Sounds abstract. So we tested it.
The test that reassured us
We took a paragraph a human had written by hand — a calm, atmospheric text about a woman on a plane with her sleeping dachshund. Unmistakably human.
Pangram: 0% AI. “Human Written,” high confidence.
That was the good news. The detector doesn’t just cry “AI” at everything. It recognizes genuine human text for what it is.
The test that worried us
Then came the moment that turned everything around. We took that exact same human paragraph and ran our own editing tool over it just once — with the instruction “make the text more literary and expressive.” Exactly what you do when you polish something.
Result: 100% AI.
A single AI polishing pass had turned genuine human text into “AI.” And that is the real insight behind almost every false positive: in most cases the detector isn’t broken at all. The text really did pass through an AI — while being “prettied up,” “smoothed out,” “made more elegant.” The very tools meant to make your text nicer leave behind the fingerprint the detector reacts to.
What did NOT work
Before we found the solution, we ran into several dead ends. For honesty’s sake:
Use a different model. The obvious idea: if one model has a recognizable fingerprint, let a different one do the rewriting. We tried four different models. Result: all four at 100%. The fingerprint isn’t model-specific, it’s AI-general. Switching models does exactly nothing.
Sprinkle in “human messiness.” The classic of every humanizer trick: filler words, colloquialisms, deliberately clumsy sentences. It didn’t work — and sometimes made the score even worse. The reason is bitterly logical: when an AI sprinkles in “human errors,” that itself is a pattern again. You can’t wash away AI detectability with even more AI.
What works: leave out instead of add
The breakthrough came when we stopped adding things.
The telltale AI sound sits in the literary polish: ornate metaphors, “as if” similes, long nested sentences, and always the most effective, most fitting word. That “too beautiful” is exactly the signal.
So the solution is radically simple: shorten and simplify. Cut the metaphors. Drop the “as if” comparisons. Replace the elevated word with the plain one (“drift” becomes “lie,” “slumber” becomes “sleep”). Break long arcs into short sentences. Say only what happens.
In our tests, exactly this reliably brought the flipped text back to 0% — human, pass after pass. The content stayed intact; the text just got plainer and terser.
The honest limit
One thing we don’t want to be misunderstood: this is not an “AI-cloaking machine.”
We also tried it on a long, fully AI-generated text. It stayed at 100%, no matter how hard we cut. And that’s how it should be: beneath genuine human text lies a human core that can be uncovered again. Beneath thoroughly generated text lies — nothing. The tool helps real writing that has picked up an unfair stain. It does not disguise machine text.
Better still: don’t generate too smoothly in the first place
The best repair is the one you don’t need. So we asked: can the AI write more plainly from the start, so the suspicion never arises at all?
The honest answer has two halves. If you have an AI write a text from nothing — no matter how often you put “write human” into the instruction — it stays AI to the detector. No human anchor, no effect.
But almost no one writes that way in Hermes. Usually there’s already a piece of you on the page, your own text, and the AI continues it. And that’s exactly when something happens: instruct the AI to keep writing pointedly plainly — short sentences, no metaphors, the simple word — and the continuation leans on your tone and blends into it. In our tests the whole text then stayed “human” in five out of six cases, where normal generation jumped straight to “AI.”
Out of that came a fourth building block.
Four tools that are now part of Hermes
Out of all this, four concrete features emerged:
1. The AI check. Check for yourself at any time where you stand — with a little gauge that shows how “human” or “AI-like” your text currently reads. So you can see in black and white whether there’s any need to act at all.
2. “Humanize text.” A one-click tool in the editing menu. It takes exactly the polish that looks like AI off the selected passage — shortening, simplifying, de-ornamenting — and gives you a plainer version you can accept or discard.
3. The AI-slop analysis. If you’d rather do it yourself: it shows you concrete spots that sound “too polished,” each with a plainer suggestion. You work through them one by one — accept, reject, next — and keep full control over your voice.
4. “Write human.” A skill you can add to any AI generation. It makes Hermes write plainly and naturally from the start — most effective when it continues your own text. That way, ideally, the stain never appears at all.
The real lesson
Perhaps the most important insight from this excursion isn’t a technical one, but one about writing itself: over-polishing is what sounds like a machine. Plain, concrete, unexcited prose — the kind a human writes on the first pass, before making it “beautiful” — reads as the most human. To a detector, and often to readers too.
And if a false alarm ever does hit you: now you know why — and the tools are ready.
Try it out
Write your book with a tool that thinks along with you — and that helps you when a detector gets it wrong. Start for free with Hermes 3000.