Is this AI text checker free?

Yes. Paste a text into the box at the top and press Check. There is no sign-up and no limit. The first check takes a few seconds to load; after that each one takes about a second.

Is this an AI detector?

No. Detectors guess whether a model wrote a text and are often wrong both ways. This check looks for specific phrases a human reader uses to spot generated text and shows where they are. A person who writes in corporate boilerplate gets the same findings, which is fair, because a reader will take that text for a model’s too.

Can anyone prove a text was written by AI?

No. Neither a detector nor this check gives proof. A detector gives a probability that is also wrong about people. This check shows the phrases that make a text look generated, and that is something you can argue with point by point.

Why did an AI detector flag my own writing?

Statistical detectors react to predictability. Plain words, even sentences and standard phrasing look like a model to them, and plenty of people write that way, especially in business writing or in a second language. Paste the text here to see which places read as a template.

Will my text pass an AI detector after I fix it?

Nobody can promise that. A detector runs its own statistics and its answer cannot be known in advance. Fixing the findings removes what a reader notices, and the reader is who the text is for.

Does it work on ChatGPT, Claude and Gemini output?

Yes. The check looks at the text itself and does not care which model wrote it. The phrases in the catalogue turn up in all the popular models.

Where does my text go?

Nowhere. The linter runs in your browser on Pyodide, a build of Python for the web. The page downloads it once and works offline after that.

Which languages are checked?

Russian more thoroughly, English for the common phrases. There are no rules for other languages yet, so a clean result there means nothing.

Can I run this automatically?

The same linter installs with one command and runs in a terminal or in CI. It is free and open source; the skills page has the instructions.

AI text checker

AI text checker: the phrases that make writing read as AI

Paste a text and the page highlights the phrases a reader recognises from ChatGPT or Claude: “in today’s fast-paced world”, “it’s important to note”, “game-changer”. It is free, with no sign-up. The check runs in your browser and the text is not sent anywhere.

01It is not a detector and will not give you a percentage of “written by AI”. It shows the exact phrases that make a text read that way and what to do with them.
02Russian is covered more thoroughly; English is checked for the common phrases. Texts in other languages will come out almost clean, and that will mean nothing.
03The first check downloads about 6 MB: Python and the linter itself. Every check after that is immediate.
04The rules match the linter in the skills kit, so the same text gets the same findings here and on the command line.

What it checks

Ten kinds of phrasing that give a model away

The linter has more than forty categories; here they are folded into ten groups. Every quoted example comes from the linter’s rules and is highlighted when you run a check. The English catalogue is shorter than the Russian one.

Stock phrases

The phrases a model uses to open, link and close a thought: “in today’s fast-paced world”, “it’s important to note”, “delve into”, “moreover”. They say nothing, and readers know them by heart.

What to doDelete the phrase and reread the sentence without it. It almost always gets shorter and better.

Staged contrasts

“This isn’t about X — it’s about Y”, set-ups like “The best part: …”. A model stages depth this way, inventing a view nobody held and then correcting it.

What to doSay what the thing is. Keep a denial only if someone really believes the thing you deny.

Loud words

“Game-changer”, “cutting-edge”, “revolutionary” put a verdict where a fact should be. The louder the word, the less anyone believes it.

What to doReplace the verdict with what it rests on: a number, a date, a name, an example. If there is no fact, drop the word.

Corporate language

“Leverage”, “synergy” and nouns made out of verbs make a text sound like a quarterly report.

What to doBring the verb back and use the word you would say out loud.

Clichés and borrowed authority

“Plays a crucial role”, or “studies show that” with no study named: a stock phrase or an unnamed authority stands in for the thought.

What to doSay literally what happens. If you cite research, name it or drop the claim.

Coach and therapist voice

“And that’s okay”, “you’re not alone”, “unleash your potential”. Comfort and pep talk the reader did not ask for give the template away.

What to doCut it. If the reader needs support, a concrete piece of advice gives more of it.

Hedging

“Could perhaps somewhat arguably”: four hedges in one sentence. The model is covering itself, and the reader stops knowing what you claim.

What to doKeep one qualifier where you are genuinely unsure. State the rest.

Rhythm and repetition

Sentences of the same length in a row, the same verb in neighbouring sentences, bold every few words, a line under a heading that repeats the heading. The words can be fine and the text still drones.

What to doSplit a long sentence, join two short ones, keep one short sentence for emphasis. Replace a repeated verb or rebuild the sentence.

Typography

Straight quotes where curly ones belong, and in Russian text the long dash and signs like “=” and “→” used in place of words.

What to doFix the quotes; write the signs out as words.

Traces of copying from a chat

Service marks like utm_source=chatgpt.com and invisible characters only get into a text when it is copied out of a chatbot. One is enough to give it away.

What to doDelete the mark. If a link stood there, find the real source.

How this differs from AI detectors

A detector gives a probability. This shows the phrase.

Tools like GPTZero and the detectors built into plagiarism checkers estimate how predictable a text is to a language model. A flat vocabulary and sentences of one length push the “AI” percentage up. The detector does not say where the number came from, so there is nothing to argue with.

Checks built this way are often wrong about people. In 2023 Stanford researchers ran TOEFL essays written by humans through seven detectors, and on average the detectors called more than half of them AI-generated (Liang et al., Patterns). The same year OpenAI shut down its own classifier: it caught 26% of AI-written texts and labelled human writing as AI 9% of the time.

This check answers a different question. It finds the phrases a reader trips over, shows each one in the text and says what to do about it. It does not decide who wrote the text. A generated text that has been edited well comes out clean, and a person who writes in templates gets findings. The reader cares about the result, and so does the check.

How to fix a flagged text

What to do when the check finds a lot

01Start with the hard rule breaks, marked in red. Traces of copying from a chat and staged “this isn’t X, it’s Y” lines are worth removing whatever else you do.
02Delete before you replace. A stock phrase can almost always go with nothing lost. A synonym for “plays a crucial role” keeps the same template in other words.
03Replace a verdict with a fact. Instead of “dramatically speeds up the work”, say how long it took before and how long it takes now. If there is no number, drop the verdict.
04Read it aloud. Rhythm, repetition and corporate language are easier to hear than to see. Rewrite any sentence you would not say to a person the way you would say it.
05Check again. A couple of findings in a long text happen to people too; there is no need to chase zero.

Questions

Is this AI text checker free?: Yes. Paste a text into the box at the top and press Check. There is no sign-up and no limit. The first check takes a few seconds to load; after that each one takes about a second.
Is this an AI detector?: No. Detectors guess whether a model wrote a text and are often wrong both ways. This check looks for specific phrases a human reader uses to spot generated text and shows where they are. A person who writes in corporate boilerplate gets the same findings, which is fair, because a reader will take that text for a model’s too.
Can anyone prove a text was written by AI?: No. Neither a detector nor this check gives proof. A detector gives a probability that is also wrong about people. This check shows the phrases that make a text look generated, and that is something you can argue with point by point.
Why did an AI detector flag my own writing?: Statistical detectors react to predictability. Plain words, even sentences and standard phrasing look like a model to them, and plenty of people write that way, especially in business writing or in a second language. Paste the text here to see which places read as a template.
Will my text pass an AI detector after I fix it?: Nobody can promise that. A detector runs its own statistics and its answer cannot be known in advance. Fixing the findings removes what a reader notices, and the reader is who the text is for.
Does it work on ChatGPT, Claude and Gemini output?: Yes. The check looks at the text itself and does not care which model wrote it. The phrases in the catalogue turn up in all the popular models.
Where does my text go?: Nowhere. The linter runs in your browser on Pyodide, a build of Python for the web. The page downloads it once and works offline after that.
Which languages are checked?: Russian more thoroughly, English for the common phrases. There are no rules for other languages yet, so a clean result there means nothing.
Can I run this automatically?: The same linter installs with one command and runs in a terminal or in CI. It is free and open source; the skills page has the instructions.

Finding the phrases is half the job; they still have to be rewritten. If you work in Claude Code, the skills kit has helpers that do it for posts, emails, landing pages, release notes and prose, following the same rules as this check.

About the skills kit →

If a team writes with AI

When a model writes posts, emails and customer replies in bulk, cleaning each text by hand gets expensive. The fix is a process where the prompt, the rules and this check run before anything is published. The first consultation is free and takes thirty minutes.

Book a consultation →