Claude humanizer

What a Claude humanizer should do (and what it should not)

A Claude humanizer should take text that was drafted or edited by Claude and rewrite it so it sounds closer to you. The point is not to add random synonyms or make the text messy. The point is to preserve your meaning while removing the statistical pattern that Claude may have added.

For Unmarker, that means three jobs. First, it rewrites in the author's voice rather than in a generic model voice. Second, it checks meaning sentence by sentence, including entities, numbers, dates and quotations. Third, it shows a before and after watermark score, so you can see what changed.

A good tool should also keep structure. If you paste Markdown or HTML, headings, lists, links and citations should remain usable. If the output changes a quote, breaks a reference or alters a number, that is a failure, not a useful humanisation.

A Claude humanizer should not promise results it cannot measure. Style based AI detectors are different from watermark detectors. They try to judge whether text reads like model output, while a watermark detector looks for a vendor's key based signal.

If you only need a direct removal workflow, see the Claude watermark remover. If you want to measure the signal first, use the planned Claude watermark detector.

Why Claude text is now watermarked and what that changes

On August 11, 2026, Anthropic announced that Claude output carries a statistical text watermark. It biases word choices according to a key held by Anthropic. There are no hidden characters to delete, so cleaning invisible formatting is not enough.

The watermark applies worldwide to Claude models launched on or after August 2, 2026, including the app, Claude Code, Cowork and the API. Older models are being added progressively. The trigger is the EU AI Act, Regulation (EU) 2024/1689, article 50, whose transparency obligations are in force from August 2, 2026.

This kind of watermark survives copy and paste because it is part of the wording pattern. It may also persist through light editing. Heavy paraphrase, translation or very short passages can weaken it, but that does not mean ordinary edits reliably remove it.

A detected mark means Claude touched the text, not necessarily that Claude wrote it from scratch. A human text that was proofread or translated by Claude may be flagged. That distinction matters if you use Claude as an editor rather than as the original author.

For a plain explanation of the mechanism, read how Claude's text watermark works. For the practical rewrite workflow, read how to remove the Claude watermark.

How Unmarker rewrites: voice, meaning check, before/after score

Unmarker is an independent product, not affiliated with or endorsed by Anthropic.

Unmarker is a pre-launch product by STRATINET SARL in Cannes, France. The waitlist is open for early access. The product is built around measured rewriting, not vague polishing.

You paste Claude text and, where useful, a sample of your own writing. Unmarker rewrites with a model that carries no Claude watermark. It aims to keep your structure and intent, while replacing the statistical pattern that marks Claude output.

The meaning check is not decorative. Each sentence is compared against the source for entities, numbers, quotes and claims. If a rewrite risks changing the meaning, the difference is highlighted so you can decide whether to accept, edit or regenerate that section.

The interface is planned to show a before and after watermark score and a diff. The score is about the Claude watermark signal, not a broad judgement of whether text seems artificial. The diff lets you see how much rewriting was needed and where the wording changed.

Formatting is part of the job. Unmarker is planned to preserve Markdown and HTML, including headings, lists, links and citation shaped text. That matters for articles, documentation, product pages and code related notes where structure is not optional.

If you work across models, the broader AI text watermark remover explains the same idea for vendor watermark signals. For measuring rather than rewriting, see the AI watermark detector.

Humanizer vs paraphraser vs asking Claude to rewrite itself

These approaches sound similar, but they solve different problems.

A paraphraser can be useful for quick wording changes, but that is not the same as controlled watermark removal. If it swaps words without checking meaning, it can create subtle errors. If it rewrites too lightly, the watermark signal may remain measurable.

Asking Claude to make the text more natural is also not the same thing. It may improve readability, but the output still comes from Claude. If the model is covered by Anthropic's watermarking policy, the rewritten text may continue to carry a Claude signal.

Unmarker is designed for a narrower job: rewrite with a non watermarked model, compare meaning, preserve formatting and show the before and after score. That makes the result auditable by you before you use it.

What we will not claim

We will not claim that text will pass AI detectors. Unmarker measures the Claude watermark signal, not stylometric detectors such as GPTZero, Turnitin or Originality. Those systems answer a different question, and they can produce false positives.

We also will not say that every short passage can be judged with confidence. Very short text gives less signal to measure. A single sentence, heading or fragment may not contain enough wording pattern for a useful watermark score.

We will not advise academic or professional dishonesty. If your school, employer, publisher or client requires disclosure of AI assistance, follow that rule. A rewriting tool does not remove your responsibility for the final text or for any required attribution.

We will not pretend that watermark detection proves authorship. A Claude mark means Claude touched the text. It does not prove that Claude wrote the whole document, and it does not prove intent.

We will not make claims about Anthropic's forthcoming official detection tools before they are available. Unmarker is built to measure the watermark family described publicly, but official vendor tools may have their own thresholds and procedures.

Planned pricing

Unmarker is pre-launch, so pricing may change before launch. The planned Free tier includes 3 texts per day, up to 1,500 words. A free detector is also planned so you can check whether a text carries a measurable Claude watermark signal.

The planned Pro price is $9 per month. The first 500 founders can lock $5 per month. This is intended for people who regularly revise Claude output and need formatting preservation, meaning checks, diffs and before and after scores.

The planned API price is $0.50 per 10,000 words. It is intended for workflows that need automated checks or batch rewriting, while keeping the same constraints: preserve meaning, keep structure and show measurable change.

You can join the waitlist for early access. For questions before launch, contact contact@alexandrefuchs.fr.

Questions people ask

Here are the main questions about using a Claude humanizer.

Questions people ask