Rewrite AI drafts so they read like a person wrote them. Detectors also flag plenty of human writing as machine-made, and this site fixes that too.
Paste an AI draft, see what a detector would flag, rewrite it in one click, and export a record of your own writing if anyone questions it.
Drop in an AI draft, or your own writing if a detector has already flagged it. A check reads 65 words or more.
You get a score from 0 to 100 and a plain verdict, so you know where the draft stands before anyone else runs a check on it.
One click rewrites every sentence, or take the suggestions one at a time and keep control of the wording.
Export a signed record of your drafting history. Your school can verify it without an account and without contacting us.
A real admissions paragraph, rewritten. It had seven tells before the rewrite and none after, and it still says the same thing.
Detectors score how predictable your writing is, not who wrote it. Four groups get flagged far more often than anyone else.
Second-language writers learn the textbook constructions detectors score as machine-made. Correct grammar reads as low perplexity.
For second-language writers →Formal register and precise, structured phrasing are common autistic writing styles. To a classifier they look like a language model.
Dictation, word prediction and grammar tools smooth text into even patterns. That evenness is exactly what a detector scores as machine-made.
Writers rebuilding language after TBI or stroke tend toward shorter, evenly paced sentences. Low burstiness, high false-positive risk.
These are the patterns detectors key on. Every one is named, found in your text and fixed by the rewrite.
An AI detector does not know who wrote your text. It scores how readable the writing is to a language model — and the strongest detectors do it with two models, not one. A general model and a model tuned on human writing both read your text. Human prose slides through the tuned model, because it was tuned on exactly that kind of writing. Machine text makes it stumble. The gap between the two readings is the signal that gets flagged.
general model human-tuned model
The gap is what makes a strong detector hard to fool. A writer trying to sound human only needs to satisfy one model, but a detector that compares two keeps its footing.
In today’s rapidly evolving landscape, organisations must leverage innovative solutions.
The board met twice — March and June — and lost the minutes.
Both drafts run to the same number of words on the same subject, and the flat row is the one a detector scores as machine-made.
The tuned model is calm on the connective tissue and trips on the stock opener, the hedge and the polished close — the words a human writer would not reach for.
A human sentence reads smooth to both models. A generated one trips the tuned model on the words a human writer would not reach for, and that is where the score comes from.
Nothing in the score checks whether your argument holds, whether your citations are real, or whether the grammar is correct. A wrong essay and a right one score the same if they are written the same way.
A detector sees the text in front of it and nothing else. It has no record of who typed it or how long it took, so an identical paragraph scores identically every time.
The charts above illustrate what the measurements mean. They are drawn for the explanation, not read off your text.
You cannot lower an AI detection score by swapping words for synonyms. What moves it is changing the pattern: sentence lengths that vary, the setup sentence at the top deleted, and a specific detail only you could know. Most of the advice passed around in group chats does nothing at all, and some of it makes things worse. There is more on bypassing AI detection and on making AI text undetectable.
These change the sentence rhythm, the structure or the specificity, which is what a detector measures.
Put a five-word sentence next to a thirty-word one, then run the check again. Of everything on this list, this is where the number usually moves first.
A generated draft almost always restates the question before it answers anything. Delete that sentence and start on your actual claim.
Words like "Consequently" and "Importantly" at the head of a paragraph often connect nothing. Cut the word, read the paragraph again, and it makes the same point without it.
Generated drafts run three or four paragraphs of near identical length. Merge two, split one, and let a single sentence stand alone as its own paragraph.
Most drafts carry one sentence that restates the point made just before it. Deleting it lowers the score and shortens the draft at the same time.
A date, a course code, the real title of the book, what your supervisor said in week three. No model can predict a detail it was never given.
Commonly recommended, and all three leave the underlying pattern exactly as it was.
Changing "utilize" to "use" fixes one word and leaves the rhythm, the structure and the paragraph shape exactly as they were. The number barely moves.
Spelling is not the pattern being scored, so the draft reads the same to a detector. The errors still cost you marks with the person grading it.
A round trip through another language swaps vocabulary and leaves sentence structure where it was. The grammar usually comes back worse than it went in.
These do not lower the score either, and they cost you marks or credibility with the person reading.
"Honestly" and "at the end of the day" dropped into academic prose read as costume. The structure underneath is unchanged and the draft now sounds like two different writers.
Zero-width characters and white text do not survive submission, because portals strip formatting on paste. A grader who finds one stops reading your argument and starts questioning your honesty.
Change one thing at a time. Run a check, make a single edit, run it again, and see whether the number actually moved before you touch anything else.
What to do depends on what you are writing and how long you have. An essay due tomorrow, your own work flagged by a course tool, a personal statement, a cover letter, a dissertation chapter, marketing copy, a job application. Each one below has the steps and what it costs in credits.
Check the whole draft first, so you know which sentences are flagged before you touch them. There is a longer walkthrough for checking an essay. Rewrite those sections, then read the result before you submit it. If a sentence no longer says what you meant, put your own wording back.
Run the check yourself before the conversation starts, so you walk in knowing the score and the flagged sentences. If it has already happened, read what to do when you are wrongly accused. If the flags land on sentence rhythm and stock phrasing rather than content, rewriting fixes them without changing what you said.
Generic usually means the sentences are interchangeable, not that the story is weak. The personal statement page goes through this line by line. Rewrite only the paragraphs that read as machine-made. Leave the lines carrying your own names, dates and details exactly as they are.
Paste the whole letter in one go, since a single paragraph often falls under the 65-word minimum for a check. After the rewrite, run the AI vocabulary checker for the words every other applicant is also using.
Sections written months apart often read like different writers, which the dissertation page covers at length. Check each one separately and compare the scores. Rewrite the section that sits furthest from the others instead of the whole chapter.
Rewrite the paragraphs that come back flagged, then put a real number or a named example where each hedge was. See the page on blog posts and marketing copy. Check the reading level after the rewrite rather than before, since rewriting changes it.
A detector percentage is a reading of style, not evidence about who wrote the text. How teachers detect AI writing sets out what a number can and cannot support. Paste a sample of the same student's older work, from before this course, and see what it scores. If the two scores look alike, the number is describing how that student writes.
Perplexity, burstiness, base rate and threshold are the four numbers that decide an AI detection argument — and almost nobody explains them to the person on the receiving end. Read below to understand each.
How surprised a language model is by the next word in a passage. Predictable wording scores low, unexpected wording scores high.
Textbook grammar, a memorized essay structure and careful second-language phrasing all read as predictable. A scorer reads predictable as machine-written.
The variation in sentence length and structure across a passage. Human drafts swing between short lines and long ones. Generated text holds a steadier rhythm.
Editing until every sentence runs about the same length lowers burstiness. Polishing your own work can raise your score.
Human writing that a detector reports as AI, which is the false positive problem. The reverse error, generated text passing as human, is a false negative.
The two errors land on different people. A false negative costs the tool some accuracy. A false positive costs you a grade and leaves you proving you wrote your own paper.
How much AI writing is actually in the stack being checked, before any tool runs. An accuracy claim means little without it.
If most of a class writes its own work, most flags land on students who did too, even at a low error rate. Ask how many papers the tool expects to be AI, not how accurate it says it is.
The cutoff score where a tool stops saying human and starts saying AI. Whoever runs the tool picks where it sits.
The same paragraph passes under one threshold and fails under another. A score a point over the line is a setting someone chose, not a finding about your writing.
A number from 0 to 100 estimating how machine-like a writing pattern looks, explained in full on what an AI score means. The checker on this page returns that number with one of three verdicts: reads as AI-generated, borderline, or reads as human-written.
It describes the text and nothing else. No score can see who typed the words, so a high one is a reason to look closer, not a finding of fact.
The share of your text matching sources already in a plagiarism index, which is a different thing entirely: see AI detection versus a plagiarism check. Similarity points to a document you can open and compare. An AI score points to a pattern with no document behind it.
A paper can return zero similarity and a high AI score at the same time. Neither number tells you anything about the other.
Evidence of where a document came from and how it was written, which is what the signed drafting record produces. One form is a record of your drafting session, signed with ECDSA P-256, that a third party can verify on their own.
Every other term here produces an opinion about finished text. Provenance produces a record of how the text got made, which is what an accusation actually asks for.
An AI humanizer rewrites AI-generated text so it reads like a person wrote it. It changes the things AI detectors score: repetitive sentence rhythm, stock openers like "In today's rapidly evolving landscape", stacked hedges, and words such as delve, leverage and foster. Paste your text above and the rewrite happens in one click, free.
Yes. The AI checker, the AI detector score, the tell explanations, the free tools and the signed evidence export are free with no account. The one-click rewrite is free too, with a daily allowance.
Yes, that is what it is built for. On the sample above, the risk score drops from 85 out of 100 to 8 and all 7 tells disappear. Nobody can promise a specific number on a specific detector, because detectors are retrained without notice and disagree with each other on the same paragraph. That is also why human writing gets flagged. Check your text before you submit it rather than after.
They measure how predictable your writing is to a language model. Generated text is unusually predictable: even sentence lengths, common word choices, low variation. Detectors score that pattern, not authorship. Run the checker above and you get the same kind of reading before anyone else runs one on you.
Because detectors score style, not authorship. Writing that is clear, formal and evenly paced looks machine-made to a classifier. Second-language writers, autistic writers and anyone using dictation or grammar tools get flagged most. A Stanford study found 61% of TOEFL essays by non-native speakers were falsely flagged across seven detectors.
It depends on the rules you are working under. Rewriting a draft you were allowed to generate is ordinary editing. Where AI is not permitted in submitted work, that rule covers the draft however it is edited. Read your institution's policy. If you wrote the text yourself and were flagged anyway, export a signed record of your drafting history.
Your text is never sent to an external detection service. The tell explanations and every free tool run in your browser and send nothing anywhere. Pressing Check or Humanize does send that text to our own servers, because the work needs a machine your browser does not have. It is processed to return your result, then dropped: not stored, not used for training.
Counted per network, so a phone and a laptop on the same wifi share the same 10. The count resets the next day.
Needs at least 65 words. You get a score from 0 to 100 and one of three readings: reads as AI-generated, borderline, or reads as human-written.
Needs at least 25 words. Ten credits covers about 500 words of rewriting.
Split anything longer and run it one section at a time.
All six are free with no daily cap, because they run in your browser rather than on a server we pay for.
The rewrite changes wording and sentence rhythm, so the sentences will not be the ones you typed. Facts, names, numbers and citations carry through, but read the output against your source before you submit. If a figure or a quote shifted, fix that line by hand.
Yes. Both tabs work in a phone browser, with no app to install and no login. Paste from your notes app and run the check there.
Below 65 words there is not enough text to read a pattern from, so any score would be noise. Paste the paragraphs around it and check the passage as a whole. The same goes for the rewrite below 25 words.
Yes. Grammar assistants even out sentence length and swap your phrasing for the most common version of it, which is the pattern detectors score. Run the check on the corrected draft rather than assuming a clean grammar pass reads as human.
Nothing here reports to a school, an instructor or a detection service. There is no account, no login, no card and no email, so there is nothing on file under your name. If you are working on a school laptop, that machine still keeps its own browser history.
Borderline sits between the two clear readings. Look at which tells are marked, then fix those yourself: vary sentence length, cut the stock opener, put your own example back in. Re-run and compare the two numbers.
As many times as your credits allow, at 1 credit per 50 words each pass. Each pass rewrites the output of the last one, so repeated runs drift further from what you meant. One pass plus your own edit usually reads better than three passes.
That depends on the policy you are writing under, and policies differ by course, not only by school. Check the syllabus or the assignment brief for a disclosure line, and ask your instructor in writing if there is not one. A written answer is worth having later.
Run the checker on the draft you already have, which needs 65 words. Read the tells it marks, fix the two or three worst ones yourself, then re-run. Export the signed drafting record last, so you have something to show if the submission is questioned.
Free AI checker and AI humanizer. No account, no card, no sign-up.