Claude AI detector
A Claude AI detector is more reliable when it checks for Claude's statistical watermark, rather than guessing from writing style. Unmarker is being built to test for that signal, with clear limits: a mark means Claude touched the text, not that Claude wrote every word.
Founder pricing for the first 500. No spam, one email at launch.
- August 11, 2026: Anthropic announced a statistical text watermark for Claude output.
- Applies to Claude models launched on or after August 2, 2026.
- Very short passages give uncertain watermark results.
- Free plan planned: 3 texts per day, up to 1,500 words.
- Pro planned at $9/month, with $5/month for the first 500 founders.
Paste, Rewrite or Check, Verify style.
Paste
Paste the text you want to assess. Keep the wording as close as possible to the version you received, because heavy editing, translation or paraphrase can weaken the statistical signal.
Rewrite or Check
Choose a watermark check if you only need a result. Choose rewrite if you want Unmarker to produce a new version using a model that carries no Claude watermark.
Verify style
Review the score, diff and sentence meaning checks. Treat the result as a watermark finding only, not as proof of authorship or a prediction about style-based detectors.
Style-based AI detectors vs Claude's watermark
Most AI detectors look at style. They try to answer whether a passage reads like model output, based on patterns such as fluency, predictability, sentence shape and vocabulary. That can be useful as a clue, but it is not the same as checking for a vendor watermark.
Claude's watermark is different. On August 11, 2026, Anthropic announced that Claude output carries a statistical text watermark. It biases word choices according to a key Anthropic holds. There are no hidden characters to delete.
This means the better question is not only "does this sound like Claude". The better question is "does this text carry Claude's key-based signal". A Claude watermark detector is designed around that second question.
| Check type | What it asks | Main limit |
|---|---|---|
| Style-based AI detector | Does this read like a model wrote it | Documented false positives, especially on some human texts |
| Claude watermark check | Does this carry Claude's statistical signal | Only meaningful where the watermark may have been applied and preserved |
Unmarker focuses on the watermark measurement, not stylometric AI detectors such as GPTZero, Turnitin or Originality. It makes no claim about their outputs. If you want the broader vendor context, see which AI vendors watermark text.
Unmarker is an independent product, not affiliated with or endorsed by Anthropic.
What a positive result means (and does not)
A positive result means the text appears to carry a Claude watermark signal. It does not prove that Claude wrote the whole passage. It means Claude touched the text in a way that left the statistical pattern detectable.
That distinction matters. A human may write a draft, then ask Claude to proofread, translate, shorten or polish it. If the output comes back with Claude's watermark, the mark can reflect that intervention rather than original authorship.
The watermark applies worldwide to Claude models launched on or after August 2, 2026, including app, Claude Code, Cowork and API use. Older models are being added progressively. A negative result on text from an older model is therefore not proof that Claude was never involved.
The trigger was article 50 of the EU AI Act, Regulation (EU) 2024/1689, whose transparency obligations came into force on August 2, 2026. Anthropic's detection tools have been announced as forthcoming. Unmarker is a pre-launch independent product by STRATINET SARL, based in Cannes, France.
For a plain explanation of the mechanism, read how Claude's text watermark works. The underlying method family is statistical watermarking, where model word choices are biased towards preferred sets. This is related to the green list and red list idea described by Kirchenbauer et al. in 2023.
False positives: proofread and translated texts
False positives need careful language. With style-based systems, documented false positives mean that human writing can be classified as AI-like. With Claude's watermark, a different false positive can occur: a human text can be flagged after Claude proofread or translated it.
In that case, the text is not simply machine-authored. It may be a human text that was processed by a watermarked Claude model. The honest conclusion is narrower: Claude likely touched the final wording.
This is why Unmarker avoids making authorship claims beyond the measurement. It can show a watermark score and a before and after comparison, but it should not be treated as a complete history of who wrote each idea. You still need context, drafts and normal judgement.
Translation is especially important. If a person writes a text in one language and Claude translates it, the translated output may carry the mark. Light editing may also leave enough of the pattern in place.
How to check a text
To check a text, paste the passage into the detector when early access opens. The planned free detector will report whether the Claude watermark signal appears measurable. It will also be clear when a passage is too short or too altered to support a confident result.
Use the result as a watermark finding, not as a moral judgement. If the tool reports a positive signal, the safe wording is that Claude appears to have touched the text. If the tool reports no signal, the safe wording is that the tested passage does not show a measurable Claude watermark under the tool's conditions.
Unmarker is also being built to rewrite Claude text using a model that carries no watermark. That rewrite flow checks every sentence for meaning, including entities, numbers and quotes. It keeps Markdown and HTML formatting, then shows a diff and a before and after watermark score.
If your need is removal rather than checking, use the Claude watermark remover page. If your need is a broader check across watermark types, see the AI watermark detector. Early access is available through the join the waitlist link.
Limits stated plainly
Short texts are uncertain. Statistical watermarking needs enough tokens to produce a measurable pattern. A few sentences may not contain enough signal for a stable conclusion.
Heavy paraphrase weakens the signal. Translation can weaken it too, although translated Claude output may also be watermarked if Claude produced the translated wording. Copy-paste does not remove the watermark, because it is not stored as hidden characters.
The detector is not a stylometric AI detector. It does not measure whether a text will be flagged by GPTZero, Turnitin, Originality or similar tools. It measures the watermark signal it is designed to measure.
It also does not cover every AI vendor in the same way. Gemini uses SynthID-Text, a statistical watermark deployed since 2024, and Google has a SynthID Detector portal. OpenAI has an internal text watermark as of August 2026, but it is not deployed. Llama has open weights and no enforceable text watermark.
Because vendor systems differ, the word detector can be misleading if it is not qualified. A Claude AI detector should not be treated as a universal AI authorship detector. It is a check for a particular Claude-related signal, within the known deployment dates and editing limits.
Planned pricing
Unmarker is pre-launch, with a waitlist open for early access. A free detector is planned, so you can check limited text without starting a paid plan. Launch prices may change before release.
| Plan | Planned limit or price | Intended use |
|---|---|---|
| Free | 3 texts per day, up to 1,500 words | Occasional checks |
| Pro | $9 per month, or $5 per month locked for the first 500 founders | Regular checking and rewriting |
| API | $0.50 per 10k words | Automated workflows |
The product is developed by STRATINET SARL, Cannes, France. The founder is Alexandre Fuchs. For contact, use contact@alexandrefuchs.fr.
Pricing is not a promise of final terms. The product is still being prepared, including the detector, rewrite flow, sentence meaning checks, watermark score display and diff view. The aim is to make the scope visible rather than imply more certainty than the signal can support.
Questions people ask
These are the short answers to the main questions about using Unmarker as a Claude AI detector.
Questions people ask
Can you detect if a text was written by Claude?
Unmarker is being built to detect Claude's statistical watermark, not to prove authorship. A positive result means Claude appears to have touched the text. That can include drafting, rewriting, proofreading or translation. It does not prove that Claude wrote every sentence, and short or heavily edited texts may be uncertain.
Is the Claude AI detector free?
A free detector is planned for launch. The planned Free tier is 3 texts per day, up to 1,500 words. Pro and API plans are also planned, but prices may change before launch. Early access is available through the waitlist while the product is still pre-launch.
Does it work on old Claude models?
The watermark applies to Claude models launched on or after August 2, 2026, including app, Claude Code, Cowork and API use. Older models are being added progressively. That means older Claude output may not carry the signal yet, so a negative result is not proof that Claude was never involved.
Can Claude text pass AI detectors?
We do not measure that. Unmarker measures the Claude watermark signal, not the outputs of style-based AI detectors such as GPTZero, Turnitin or Originality. It makes no claim about how those systems will classify a text. Their methods and thresholds are separate from Claude's watermark.