Paste the ChatGPT draft into the box above and press humanize. The rewriter goes through every sentence in one pass and hands back a version with the constructions from our tell registry written out of it. One run takes 12,000 characters, about 2,000 words, and costs 1 credit per 50 words. The free tier is 10 credits a day per network, with no account, no card and no email address to verify. Run the checker on what comes back for a 0-100 score and a verdict in seconds. The rest of this page covers the ChatGPT habits that show up most often and how to rewrite each one by hand.
The humanizer accepts 12,000 characters per run, roughly 2,000 words. A blog post or a standard essay clears in a single pass. A 5,000-word paper takes three runs, so split it at section breaks rather than mid-argument. The minimum is 25 words and the cost is 1 credit per 50 words.
The checker is separate and costs 1 credit per run whatever the length. It needs 65 words and returns in seconds with a 0-100 score and one of three verdicts: reads as AI-generated, borderline, or reads as human-written. Run it before and after a rewrite to see which paragraph carried the score.
Several tools alongside it are free and unlimited, and none of them spend a credit. The registry of named AI writing tells highlights the exact phrases in your own text. There is also a proofreader, a word counter, a readability checker, an AI vocabulary checker, a plagiarism and originality check, and a signed drafting record.
A ChatGPT draft rarely gets caught on one exotic word. It gets caught on a cluster of ordinary constructions that arrive together in a short span. The registry names them in six groups: structure, openers, transitions, lexicon, hedges and cadence. Everything below is a real entry, and each one is fixable in a single edit.
These are the groups you will hit first in chat output:
Swap every flagged word and a draft can still read as machine-made, because part of the giveaway is the shape of the sentences rather than the vocabulary. Chat output tends to hold one sentence length for a whole paragraph, and a flat length distribution is what the registry's cadence detectors look at.
The tell registry flags a passage when sentence lengths vary by less than 0.32 as a coefficient of variation, measured across the passage. Human drafts usually land above 0.45. The fix is mechanical. Cut one sentence into two, then let the next run long. A four-word sentence in the middle of a paragraph does more for that number than any synonym.
Two other rhythm tells are worth a pass. Em dashes are flagged when a passage has three or more of them and at least 1.6 per 100 words, and commas and full stops do the same work. Three sentences in a row opening on the same word is the other, and the repair is to start one of them on the object.
Every fix below follows one of four rules. Keep the positive half of a pivot. Cut a connective and check whether the paragraph still holds, because it almost always does. Replace a rating adjective with the thing at stake. Put a specific where a quantity word sat.
Check first, so you have a baseline number to beat. Then paste the same text into the tell registry, which highlights each phrase in place with the reason it was flagged. Some flagged phrases are right in your context and you should keep those. Then humanize, re-check, and read the result aloud once.
A rewrite does not check facts. Any name, date, number or citation that came out of the chat window still needs checking against a source, and models produce citations that look right and do not exist. One specific that only you know also does more for a paragraph than any amount of rephrasing, so put yours in while you are there.
On academic use the site takes one line. Rewriting a draft you were allowed to generate is ordinary editing. Where AI is not permitted, that rule covers the draft however it is edited afterwards. Read the policy that applies to you.
A detection score measures how machine-made a piece of text reads. It does not measure who wrote it, because the only evidence in the file is the prose. Careful formal human writing collects the same tells.
That matters most for writers who get flagged without touching a chat window. Liang et al. at Stanford HAI in 2023 ran TOEFL essays by non-native English speakers through seven detectors, and 61% were falsely flagged as AI. Formal register and a narrow idiom range look, to a detector, like generated text.
If you wrote the piece yourself and expect to be questioned, the signed drafting record is free and unlimited. It records the session as you write, including timestamps, how much you cut and whether anything large arrived in a single paste, then exports a signed file that anyone can check at verify a report.
We make no claim about what any other detector will output. What the humanizer does is rewrite every sentence and remove the constructions in our tell registry. Run our checker on the result and read the score yourself.
12,000 characters per run, about 2,000 words. For anything longer, split at a section break and run the parts separately so each run works on one argument instead of a whole paper.
No. The free tier is 10 credits a day per network with no login, no card and no email address to verify. The humanizer costs 1 credit per 50 words, the checker 1 credit per run at any length.
Yes. The registry matches on phrasing rather than on the source, so the same entries fire on output from Claude, Gemini, Copilot or anything else. Chat models share most of these habits.
That happens, and it is not evidence that you did anything. These tells are common in careful formal writing and in anything edited toward a house style. Use the tell list to see which phrases drive the score.
The humanizer needs 25 words and the checker needs 65. Below that there is not enough text to read.