AI Text Detector

Paste text or a URL to estimate the probability that it was AI-generated and see which of 12 statistical metrics point that way. English, Ukrainian, Russian, Polish and 5 more languages.

Free, no signup. Up to 50,000 characters; most reliable from 150–200 words.

Enter a page URL — text will be automatically extracted for analysis

Language

AI text detector: what it checks

A free online AI text detector: it computes 12 statistical signals that usually separate AI-generated text from human writing, from uniform sentence length to stock AI connectors, and combines them into a 0–100% probability. It is an estimate, not proof: the result shows what the text resembles and explains why.

AI probability from 0 to 100% and a verdict: human, mixed or AI

12 metrics: entropy, burstiness, vocabulary, Zipf's law, stylometry, AI phrases, punctuation, coherence, grammar, sentence openings, readability consistency, rhythm

Explanations: which signals point to AI and which point to a human

Automatic language detection; AI-phrase dictionaries for English, Ukrainian, Russian and Polish, statistics for 5 more languages

Check pasted text or a page by URL

Free, no signup, up to 50,000 characters

How to Use

Paste the text (ideally 150+ words) or switch to a page URL
Keep automatic language detection or pick the language and click Check
Review the overall score, confidence and metrics with explanations; for a sentence-level breakdown open AI Detector PRO

FAQ

Paste the text or enter a page URL and click Check. No signup is needed; the limit is 50,000 characters per check and 20 checks per minute.

No detector can prove authorship, and we do not claim an accuracy percentage. The result is a probability estimate: it is steadier on longer texts (150–200+ words), and the Confidence field tells you how much weight it can bear.

False positives come from formal or bureaucratic style, stock connectors ("moreover", "in conclusion"), sentences of uniform length, non-native writing and translated text. Check which metrics are highlighted: they are the reasons behind the score.

It does not name the model. The detector looks for statistical traits of machine-written text shared by different models. AI text that a person has substantially edited can score low.

Language is detected automatically. English, Ukrainian, Russian and Polish have a dictionary of typical AI phrases; German, French, Spanish, Italian and Portuguese are scored on the statistical metrics only.

This page gives a fast overall score for the whole text. PRO breaks the text down sentence by sentence, highlights AI-like passages and adds a metric radar and text statistics.

How the AI text detector works

An AI detector does not "recognise" ChatGPT or look your text up in a database of model outputs. It measures how statistically similar the text is to machine writing. Language models write smoothly and predictably: they pick the most likely words, build sentences of roughly equal length and lean on connectors such as "moreover", "furthermore" and "it is important to note". People write more unevenly: a short line, then a long sentence with a digression, a repeat, an unexpected word.

Our detector computes 12 such signals and combines them into a single probability. The heaviest is AI-typical words and phrases, followed by burstiness (how much sentence length varies). The rest add context:

Entropy and vocabulary: how predictable the word choice is and how often words repeat.
Zipf's law: whether word frequencies follow a natural distribution. It is noisy on short texts, so it carries the least weight.
Stylometry, punctuation, grammar: an even formal style, monotonous punctuation, suspiciously flawless formatting.
Sentence openings, rhythm, coherence, readability consistency: repeated openers, a flat rhythm, overly smooth transitions, the same difficulty from first paragraph to last.

Everything is computed on our server with plain statistics; the text is not sent to third-party AI services. There is no neural network inside, which is why the result is explainable: every score comes with the metrics behind it.

How to read the score

Score	Meaning
under 35%	Statistically the text looks human-written.
35–65%	Mixed signals: an AI draft edited by a person, or human writing in a very formal style.
65% and above	Many traits of machine-written text at once.

Next to the percentage you will see confidence. It grows with text length (up to about 200 words) and with how much the metrics agree. Low confidence means "don't draw conclusions yet": add more text or look at the individual metrics. Below the score, explanations list what points to AI (for example, uniform sentence length) and what points to a human (varied openers, a lively rhythm). To see which sentences push the score up, open AI Detector PRO.

Where AI detectors go wrong

Short texts. Three to five sentences are too little for statistics. Results are steadier from 150–200 words.
Formal and bureaucratic writing. Reports, legal text, essays and manuals are even and "correct", which is exactly how a model writes.
Non-native writers. People writing in a second language tend to use common words and standard connectors, and get flagged more often.
Translated text. Machine and careful human translation alike flatten rhythm and vocabulary.

It also fails the other way: AI text that a person has substantially rewritten (reordered, added their own examples, split long sentences) can score low.

What to do if your own text is flagged

Look at which metrics are high. A high "AI phrases" score means stock connectors to replace; high burstiness means sentences that need more varied length.
Add specifics: numbers from your own experience, names, dates, examples. Models write in generalities; people write about particular things.
Check readability and grammar: an overly even text needs livening up, not more complexity.
If a model really did write the draft, remove its traces: em dashes and markdown, invisible characters, then rewrite it in your own words or with the humanizer, and proofread the result.
If the text is yours and you may need to defend it, keep the edit history (Google Docs or Word versions). It is stronger evidence than any percentage.

On hidden marks: this detector judges style; it does not look for invisible characters or swapped letters. Use the text watermark detector for that.

Languages

The language is detected automatically, or you can pick it yourself. English, Ukrainian, Russian and Polish have a dictionary of typical AI phrases. German, French, Spanish, Italian and Portuguese are scored on the statistical metrics only, so results there are coarser.

Ethics: a score is not a verdict

A detector result is a reason to look closer, not proof. Do not use it as the sole basis for accusing a student, writer or employee: formal style, a second language or translation can produce a high score with no AI involved. Talk to the author, ask for drafts and edit history, compare with their earlier work. For editors and SEO teams, the better question is not "did AI write this?" but "is it useful, accurate and grounded in real experience?"