Why I Stopped Trusting '% AI' Scores — and Built a Signals Checker Instead
I used to paste text into an AI detector, look at the big percentage, and treat it as the answer. Then I started reading what those numbers actually mean — and I stopped trusting them. Try the tool this article is about
I used to paste text into an AI detector, look at the big percentage, and treat it as the answer. Then I started reading what those numbers actually mean — and I stopped trusting them.
Try the tool this article is about here: Loona AI Detector
The percentage looked certain. It was not.
OpenAI shut down its own AI classifier because of its low accuracy. That should have been a warning for the whole category: even the company behind the most widely used AI models could not build a detector it was willing to stand behind.
Then there is the fairness problem. Stanford research found detectors flagged 61% of essays by non-native English writers as AI-generated, while the rate for native writers was near-zero. A score that treats fluent, careful writing in a second language as suspicious is not a measurement. It is a guess with consequences — for students, job applicants, and freelancers whose work gets questioned because of how they write.
A single number hides all of that. It does not tell you what it noticed, how sure it is, or what to do next. It just looks confident.
So I built a signals checker instead
That is why Loona AI Detector never prints a percentage or a verdict. There is no "87% AI" and no "human" stamp. Instead, it checks 6 signals in your text:
- Sentence-length rhythm
- Word variety
- Repeated phrases
- Sentence-length patterns
- Punctuation habits
- Common AI phrases
For each signal, it shows you what it found and explains it in plain language, so you can judge the text yourself. Taken together, the signals give you a band — mostly human-like, mixed, or mostly machine-like patterns — rather than a fake precise score. A band is honest about what a browser tool can actually tell you.
Private by design
The other thing I wanted was privacy. Loona AI Detector runs 100% in the browser. Your text never leaves your device, so there is nothing to upload, store, or leak. It is free, with unlimited length and no sign-up.
That matters when the text is a draft, a client document, or an essay you do not want sitting on someone else's server. You paste it in, read the signals, and close the tab. Nothing follows you home.
How I use it
I use it as a second pair of eyes, not a judge. If several signals point the same way, I read the passage again and ask whether it sounds like me. Sometimes a repeated phrase is just a habit I should vary. Sometimes a stretch of even, similar-length sentences is a paragraph I rushed. The signals are useful because they point at something I can actually fix, instead of handing me a number I can only worry about.
A detector should help a writer improve a draft, not label them. That is a much smaller promise — and it is one a simple tool can keep.
You can try it here: Loona AI Detector
What would make you trust an AI detector — a score, or an explanation you can check for yourself? Let me know in the comments.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.