How we actually
detect AI text
No black box magic. Here's what we learned from months of research.
I spent weeks reading academic papers on AI detection. What I found surprised me - most commercial detectors use just one or two signals. Run the same text twice, get different scores. That's not science, that's guessing.
The truth is, there's no single "AI fingerprint." You need to look at multiple things at once. That's what we do.
We combine several signals into one score. I won't give away all the details (competitors would love that), but here's the general idea:
Sentence rhythm
Humans write with natural variation. Short sentence. Then a longer one with more detail. AI tends to be more uniform - similar lengths, similar structure. We measure how much the writing "breathes."
Word choices
AI has favorite words. "Leverage," "utilize," "crucial," "comprehensive" - these pop up way more often than in human writing. We track vocabulary patterns that give AI away.
Predictability
AI works by predicting the next word. This makes AI text weirdly... expected. Humans surprise you more. We have ways to measure this "surprise factor."
Hidden repetition
AI loves certain phrase patterns. "In order to," "it is important to note," "when it comes to" - the same structures over and over. You don't notice it reading, but the math catches it.
Here's where most detectors mess up: a news article full of quotes can look "AI-like." Technical docs follow strict patterns. Legal text is formal by nature.
Our system understands context. We trained on thousands of real examples - all kinds of writing styles. A quote-heavy article won't get falsely flagged just because it follows journalistic standards.
Most detectors only work well in English. We built ours to read many writing systems, and we measure it in each language separately: English is where it is strongest, and in every language we show you what the score is worth instead of pretending it is equally sure everywhere.
Pallas is a detection model we train and update ourselves rather than license from anyone. It learns from human writing and machine writing in the languages our users actually write in, and it is retrained as new AI models appear, because each one writes a little differently.
We measure how well it separates the two in every language, and we tell you when it cannot be trusted. A score that pretends to certainty is worse than no score. In the 25 languages we have measured so far, the score you see is Pallas's and the result says so; elsewhere our earlier detector still answers while we measure, and it tells you when it is on thin ice.
Penelope, our rewriting model
Pallas judges; Penelope writes. She is a rewriting model we train ourselves on one thing only: turning machine-sounding text back into the way people wrote before AI writing existed. She is in training now. She reaches the product language by language, each one measured on texts she has never seen, and your result will say which version rewrote it. Until then, rewrites run on our current engine and are checked by Pallas.
Pallas, version by version
Each new version is judged on writing it has never seen before it may replace the previous one. If it does not win, it does not ship. We say what changed and what it covers; we do not say what it looks at, because a detector that explains its tells is a detector that teaches evasion.
| Version | Date | What changed |
|---|---|---|
| 1.3 | 14 Sep 2026 | Same 33 languages. Learned to recognise the output of our own rewriting model alongside more genuine casual writing, so everyday human text is flagged less often: on writing it had never seen, 99 of 100 human texts read as human at the same catch rate as before. Live since 14 September. |
| 1.2 | 13 Sep 2026 | Trained in 18 languages, measured in 33, live in those since 13 September. Each measured language has its own decision line, set so that human writing is flagged no more than one time in twenty. Added: Arabic, Vietnamese, Russian, Dutch, Swahili, Ukrainian, Kazakh, Mongolian, Kannada, Hungarian, Bulgarian, Romanian. |
| 1.1 | 13 Sep 2026 | Same 13 languages; machine writing now drawn from four model families instead of two. Better at recognising human writing on texts it had never seen. |
| 1.0 | 13 Sep 2026 | First version. 13 languages: Tagalog, English, Albanian, Malay, Persian, Urdu, Indonesian, Serbian, Bengali, Nepali, Afrikaans, Sinhala, Croatian. Human writing from before AI writing existed; machine writing from two model families and from humanised rewrites of it. |
AI models keep getting better. GPT-4 writes differently than GPT-3. Claude writes differently than both. So we keep testing, keep learning, keep updating.
When AI gets smarter, we get smarter too. That's the deal.
Other tools swap words with synonyms or add invisible characters to trick detectors. That's lazy. And it doesn't really work across different detection tools.
We actually rewrite the content. Keep the meaning, change how it's expressed. Add natural variation, conversational flow, the little imperfections that make text feel human.
I won't share exactly how (proprietary and all that), but the philosophy is simple: real transformation, not cheap tricks.
Want to see it in action?
Detection is free. No credit card needed.
