A few weeks ago I wanted to understand what AI text detectors actually measure. The setup was simple enough. Claude had drafted me a new bio for my personal site, based on a transcript of me explaining my background. It was accurate, the voice felt close to mine, and I was ready to ship it. But before I did, I ran it through Sapling, which is the AI detector some publishers use, and it came back at 96 out of 100. Flagged as AI. Which, to be fair, it technically was. The bio was Claude's words arranged from my facts. I started wondering what it would take for that kind of text to read as "human" to a detector, and whether the tools selling themselves as "humanizers" actually did the job. So I ran some tests.
Finding 1: detectors detect style, not AI
The first thing that broke my assumptions was a test that Claude and I ran on two writing samples. Both were 100% written by humans. One was a NIST government standards document, a formal technical text. The other was a column by Brian Krebs, the security journalist who writes at krebsonsecurity.com.
The detector wasn't wrong exactly. It just wasn't measuring what I thought it was. NIST reads like a textbook: uniform sentences, structured arguments, formal register throughout. Krebs reads like a guy at a whiteboard: short next to long, fragments, conversational voice held all the way through. The detector couldn't tell who wrote either one. It just knew that one sounded official and the other sounded like a person talking.
So the thing I'd been calling "AI detection" is actually style detection. A formal, well-structured text written by a human still trips it. An AI-drafted piece written in the right register shouldn't.
Finding 2: humanizers move the score by breaking the facts
The second test was the humanizers. I ran the Claude-drafted bio through Undetectable.ai three times with different settings. The Sapling score dropped every time, which is what the tool is supposed to do. The problem was what came back in the output.
One version reported that I was 38 years old. I'm not. Another said I was "French by choice", when I actually grew up in Brazil. A third described Carbono Zero, my first company, as "an online natural products retailer", when it was a zero-emissions courier service. Another decided I was a "keen surfer" who also played tennis, neither of which I do. And the Cyberhaven section got quietly inverted: the original bio said I own the web platform there, and the humanizer rewrote it as "we're launching our web platform", which is the exact opposite of the truth.
The facts were scrambled. The detection score dropped every time. The words that came back were no longer true.
I tried Grammarly's humanizer next, on the work-history section of the bio. Grammarly didn't invent anything, but it also didn't move the Sapling score at all. Same number before and after, with a comma swap here and a word swap there. Clean, and useless.
Finding 3: you can fix the score without touching the facts
The third test was the one that actually changed how I think about all of this. Claude and I took a technical paragraph that scored 0.9998 on Sapling, which is nearly certain AI, and rewrote it in a journalist voice. Short sentences next to long ones. Fragments for emphasis. One register held all the way through. No facts changed. The new version scored 0.0018.
The score dropped by about a factor of five hundred, and the facts were identical. The only thing that changed was the sentence rhythm and the register. Which lines up exactly with what the NIST/Krebs test had already shown.
What I do now
I stopped thinking about humanizers as tools for rewriting content. I still use them sometimes, but only to measure. When something I wrote reads as AI, it's almost always a sign that my prose has slipped into template mode, and the score tells me where it happened.
For the fix, I use Claude again, but with a short page of rules I wrote down after this testing. Every number stays exact. Every name stays exact. Every claim has to be traceable back to the source. Don't add superlatives. Don't invent characterizations. When in doubt, preserve more and elaborate less. Then I ask for a rewrite in a journalist voice, and I read every line back against the original by hand.
The bio that started at 96 on Sapling now scores 0. Every fact in it is still true. Including the detail that most of it, including the research in this post, was written by Claude.