How to tell if a text was written by AI
You cannot tell for sure from the text alone. You can notice habits that machine text often has, and treat them as a reason to reread, never as proof.
AI text usually gives itself away through habits, not mistakes. Here is what to look at, and how not to jump to conclusions.
A rhythm that is too even
People write unevenly: a short sentence, a long one, a short one again. AI sentences are often almost the same length and its paragraphs come out the same size, as if cut with a ruler.
This is the strongest signal in our own analysis. The checker measures how much sentence length and paragraph length vary from one sentence or paragraph to the next. The lower the variation, the more points the text collects. Very even text with few real sentences or few paragraphs gets only partial credit, because a short text says little either way.
Stock phrases and connectors
Phrases such as "it is worth noting", "plays a crucial role", "in today's world" and "in conclusion" show up in AI text far more often than in natural speech. So does a chain of "moreover", "furthermore" and "additionally" in nearly every sentence.
The checker keeps a list of such phrases in English, Russian, Ukrainian and Portuguese, sorted into tiers by how typical each one is. A single phrase means little. A text where they pile up is worth a second look.
General words instead of specifics
Human writing has names, numbers, places and things that happened. Machine writing leans on "many studies show" and "this matters to everyone": smooth, but nothing to hold on to.
A quick test: pick any paragraph and ask what you could check in it. If the answer is nothing, the paragraph is probably padding, whoever wrote it.
Perfect smoothness
No detours, no doubts, no opinion of its own. Every paragraph opens an idea and closes it with a tidy conclusion. Human writing is usually a bit rougher than that.
Invisible traces
Sometimes a copied text carries characters you cannot see: zero-width spaces, soft hyphens, odd quotes. They are not a sign of AI on their own, since word processors and websites add them too, but they show that the text travelled through something. We describe them in invisible characters in copied text.
An important caveat
None of these signs proves authorship. News, official documents and academic papers written by people are also even and smooth. In our own measurements, neatly edited human news scored high noticeably more often than casual writing, and on a large corpus the current analysis catches only a modest share of AI texts when false alarms are held to 1%. Modern models pass more easily than the early chatbots did. The reasons are in why detectors get it wrong.
Length matters too. Under roughly 150 words the result is marked as unreliable, and between 80 and 300 words it is close to a coin toss.
What to do next
Paste the text into the checker: it highlights places with stock phrases and too even a rhythm, and lists the evidence with line numbers. Click a highlight and the editor opens at that spot with a suggestion for what to use instead.
If the text is someone else's and the stakes are real, do not decide on the score. Follow the steps in the fair checklist for teachers and editors. To see how detection works in general, read how AI text detection works today.