Most people think beating AI detectors is about better prompts. Add some typos, vary your sentence length, sprinkle in slang, and you’re in the clear. I used to believe that too. Then I saw a post that flipped the whole idea on its head.
This LinkedIn creator ran a simple experiment. It shows how detectors really work, and why chasing a “100% human-written” score is mostly the wrong goal.
🧪 The Kitchen Word Test
The author started with something almost silly. They asked Claude for the most normal, random list possible: 50 common kitchen words, starting with “chicken, rice, salt.”
No style. No clever wording. Just a plain list of words anyone could write.
Then they pasted that list into Pangram, a free AI detection tool. The result? 100% AI-generated.
They asked Claude for another list and tested it again. Same verdict: 100% AI-generated.
That’s what got me. A list of kitchen words shouldn’t have a “voice.” Yet the detector still flagged it with full confidence. So the idea that you can prompt your way past detection with a few style tweaks doesn’t hold up. Something deeper is going on.
📚 The Wikipedia Workaround
The original poster didn’t stop there. They kept testing and turned what they learned into a prompt. And it worked: Pangram rated the output 100% human-made.
The trick is surprisingly simple. The prompt restricts Claude to only pick sentences from Wikipedia. Since Wikipedia is written by humans, the detector sees human text.
Claude isn’t really “writing” anymore. It’s selecting and stitching together sentences people already wrote. The detector is checking the words themselves, so it reads them as human.
⚠️ Why You Shouldn’t Copy-Paste It
Here’s where the post gets honest, and where it actually argues against its own trick. The author added a clear disclaimer, and it’s worth reading slowly:
- It’s annoying to write that way. Limiting yourself to existing Wikipedia sentences makes natural, flowing writing really hard.
- It’s basically plagiarism. The output is a blatant copy-and-paste of whatever source material you used.
So the creator says straight out that this is not the conclusion. The only reliable way they found to “bypass” the detector was to stop generating new text at all. That tells you a lot about where the line really sits.
🤔 The Better Question to Ask
Instead of asking “How do I avoid detection?”, the expert pushes us to ask something more useful:
- When does AI detection actually matter?
- When is it totally fine for your work to be flagged as AI?
A brainstorm doc, an internal summary, or a list of grocery words probably doesn’t need a human stamp. A personal essay, a school assignment, or a piece built on your own expertise is a different story.
“Writing is thinking. You can explore and research with AI. Sometimes write with AI. But human-written will always be human-written. No way around it.”
I think that line is the real lesson here. Detectors aren’t the enemy, and tricks to beat them miss the point. If you want writing that’s truly yours, the thinking has to be yours too.
💡 How to Apply This Today
Based on the author’s findings, here’s a practical way to use AI without obsessing over detection scores:
- Use AI for research, outlines, and exploring angles you hadn’t considered.
- Decide up front whether a piece truly needs to be human-written.
- For work that matters, write the core ideas and arguments yourself.
- Use AI as an editor or sparring partner, not a ghostwriter.
- Skip detector-gaming tricks. They often lead to clunky or copied text.
This post changed how I think about the whole “AI vs. human” debate. Check out the full LinkedIn post to see the original experiment and join the conversation in the comments.