Paste the draft that
still sounds like Claude.

Most likely that you write in a regular, even style, which is the single thing the detector weighs most heavily and the thing it is most often wrong about. Short paragraphs are the known weak spot — news-style and punchy marketing copy score far too high, and genuinely human-written news articles have scored around 99%. Formal and academic registers push it up too.

So: a high score is a probability, not a finding, and it is not calibrated — 90% is a raw model output, not a claim that 90 of 100 such texts were machine-written. If someone is using a number like this as evidence about you, that is a misuse of it, and we would say so in writing. What the reading is good for is your own judgement about your own draft before you send it.

Because it is a real analysis rather than a canned animation, and it costs us something each time. One per visitor, no account, nothing stored — which is also why this page does not spend it for you: the box above carries your text to the demo rather than running anything here, so your free run is still yours when you arrive. Signing up lifts the cap and brings 1,000 Humanizer words with it, with no card.

No, and neither does anything else you can reach. Since 11 August 2026 Anthropic embeds an invisible statistical mark in Claude's output, but reading it needs a key only Anthropic holds — there is no public detector and no third party can build one. What this page reads is the statistical signature of the writing, which is a different and much older question with a public answer.

The full picture is on our Claude watermark page. Short version: the mark rides on which words were chosen, Anthropic says a real rewrite removes the signal, and we will not put a number on that because there is nothing public to measure it against.

We don't know, and we won't claim otherwise. We grade against our own detector — the one you just used. Text that clears our 35% bar has not been shown to clear Turnitin, GPTZero, Originality.ai or any other commercial detector, and every mention of clearing a bar on this page means ours.

No rewriting service can promise you a third party's verdict. Anyone who does is guessing.

It keeps your claims, your figures, your names and the shape of the argument, and changes sentence rhythm, hedging and the phrasing a reader recognises as AI. Names and figures are swapped for placeholders before the rewrite and restored afterwards, so the real ones are never part of what gets rewritten; if a figure comes back altered, that paragraph is discarded and your original kept.

It is a model with this detector acting as its editor — not a thesaurus, and not a person. There are no human linguists in the loop and we would rather say so than sell you one. About 50 seconds for a typical five-paragraph document, two minutes at the outside. Formal writing is the hard case: roughly three in ten formal documents run out of time without clearing the bar.

For one paragraph, you probably should. The difference is that a chatbot rewrites once and hands you the result with no idea whether it worked — it cannot mark its own homework. Here every candidate is scored, and one replaces your text only if it reads measurably less machine-like and clears about a dozen quality checks: nothing truncated, no leaked model chatter, none of roughly fifty banned AI clichés.

There is also this: asking a model to improve your writing is how the output gets marked in the first place. Anthropic states plainly that using Claude to proofread your own work produces watermarked text.

The demo is free and needs no account. Signing up grants 1,000 Humanizer words, once, with no card. After that it is $9.90 a month or $99 a year — the same plan billed differently: unlimited detections, 5,000 Humanizer words a month, uploads and exportable reports on both. Run out of words and you can buy more without changing plan. Cancel whenever.

Text is processed as it arrives and is not kept, never used for training, and never sold or shared. Detection runs inside our own service, so scored text does not go to an outside company at all. Encrypted in transit and at rest. The service is operated by Girema Holdings Limited, registered in Limassol, Cyprus, so it sits under EU data protection law — you can ask for a copy of your data, a correction or its deletion at any time. Full detail in the privacy policy.