Loona AI Detector - Signals, Not a Verdict
Wondering whether a piece of writing looks machine-made? Most AI detectors hand you one confident percentage and keep their method secret. This one does the opposite: it shows its working, signal by signal. And every measurement happens in your browser - your text is never uploaded, stored or seen by us.
An honest warning - read this before trusting any AI detector
No tool can prove who wrote a text. AI detection is educated guesswork about writing style, and even the biggest names in this field get it wrong - in both directions. Treat every result on this page as signals to think about, never as evidence.
OpenAI shut down its own AI text classifier in July 2023 because of its low rate of accuracy. If the company that builds the most famous AI writer could not make detection reliable, no website's percentage deserves your blind trust - including anything you read here.
Detectors are hardest on non-native writers. In a Stanford study, popular AI detectors wrongly flagged 61% of essays written by non-native English speakers as AI-generated, while almost never flagging essays by native speakers. Simple, correct, careful English is not evidence of AI - it is often just good learning.
So please: never use this detector - or any detector - as proof, and never use a score to punish a student, reject a writer or accuse anyone of cheating. Use the signals to improve writing and to start fair conversations. That is all they are honestly good for.
How to use the Loona AI Detector
- Paste the text you want to check - 50 words is the minimum, and 150 words or more gives steadier signals.
- Click Check text and read the overall pattern: mostly human-like, mixed, or mostly machine-like.
- Read each signal to see exactly what was measured, then look at the individual sentences that stood out, with the reason each one was flagged.
How it works
Big AI detectors run trained AI models on their servers and return a single probability. This detector is built differently: it measures six statistical signals that those models also rely on, and shows you every one of them with its actual numbers.
Sentence rhythm (burstiness) measures how much your sentence lengths vary. Vocabulary diversity counts how many different words you use. Repeated phrases finds three-word phrases that come back again and again. Typical lengths compares your average sentence and word length with the textbook middle that generators cluster around. Word predictability measures how much of the text is built from the most common everyday words. Paragraph balance checks whether your paragraphs are suspiciously even in size.
The six signals are combined into one overall pattern - mostly human-like, mixed, or mostly machine-like - and the sentences that look most machine-like are listed with their reasons. There is no AI model and no server: the whole check is arithmetic on your own device, which is why it is free, instant and completely private.
Worked example
Paste a 200-word essay: you might see varied sentence rhythm and rich vocabulary (human-like signals) alongside one paragraph where five sentences in a row are nearly the same length and built from very common words (a machine-like signal). The result would read "mixed patterns" - and the sentence list shows you exactly where to look, instead of hiding everything behind one number.
Frequently Asked Questions
Is this AI detector really free?
Yes - completely free, with unlimited checks, no sign-up and no word limit imposed by us. The check runs on your own device, so it costs us nothing per check either.
Can this tool prove a text was written by AI?
No - and no other tool can either. This detector measures writing-style signals, and style is not proof of authorship. Formal, simple or template-based human writing can look machine-like, and edited AI writing can look human. Never use any detector score as the sole basis for accusing someone.
Why did my own writing show machine-like patterns?
Usually because it is formal, simple or written to a template - or because English is not your first language. A Stanford study found detectors wrongly flagged 61% of non-native writers' essays. A machine-like result on your own honest writing is a known weakness of all detectors, not evidence against you.
How is this different from GPTZero or Pangram?
Those tools run trained AI models on their servers, ask you to upload your text, and return a probability. This detector runs entirely in your browser, uploads nothing, and shows you the six measurable signals behind the reading instead of one verdict. For an important case, cross-check with more than one tool - and with a human who knows the writer.
Is my text uploaded or stored?
No. The checking happens entirely in your browser with JavaScript. Your text never leaves your device, and closing the tab erases everything.
How much text should I paste?
At least 50 words - below that, patterns are mostly noise and the detector will say so instead of guessing. Around 150 words or more gives the steadiest signals. Even long texts, though, can only ever give you signals, not certainty.
Related Text Tools
Word Counter
Words, characters, sentences, paragraphs and reading time - counted live as you type.
Word Frequency
A ranked table of your most-used words, with counts and percentages.
Reading Level Calculator
Flesch Reading Ease and grade-level scores, with a plain-English verdict.
Text Summarizer
The most important sentences of any long text, picked word for word - nothing rewritten, nothing uploaded.