AI watermark detector for text

What an AI watermark detector can and cannot see

An AI watermark detector looks for signals that may have been added when a model generated or processed text. It can help you separate a vendor watermark from ordinary AI style detection, but it cannot prove authorship on its own.

Some marks are statistical. They change word choices according to a key held by the vendor. Other marks are technical, such as invisible characters, non-breaking spaces, unusual hyphens or formatting debris. Metadata may also exist around a file, but it is not the same as a watermark inside the text.

Unmarker measures watermark signals, not stylometric AI detectors such as GPTZero, Turnitin or Originality. It makes no claim about passing them. If you need a wider overview of vendor policies, see which AI vendors watermark text.

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

The three kinds of marks: statistical, invisible characters, metadata

Statistical watermarks are the most important class for current text watermarking. The method family includes the green list and red list bias described by Kirchenbauer et al. in 2023. In simple terms, the model is nudged towards some word choices and away from others.

This kind of mark is not a hidden character that you can delete. It is a pattern across the text. It may survive copy-paste and light editing, but it becomes weaker when the passage is very short, heavily paraphrased or translated.

Invisible characters are different. They include zero-width characters, non-breaking spaces, odd hyphens and other formatting artefacts. These are often what people mean when they talk about a ChatGPT watermark, although OpenAI has not deployed its internal text watermark as of August 2026. If this is your concern, the related ChatGPT watermark remover page explains the cleaner case.

Metadata is separate again. It may describe a file, a platform export or an editing history, but it is not necessarily present in plain copied text. Unmarker focuses first on the text you paste, because that is what normally survives in email, documents, websites and CMS fields.

Which vendors mark text in August 2026

The status is not the same across vendors. Some have deployed statistical text watermarking, some have announced detection tools, and some have no enforceable text watermark in open weight releases.

For Claude-specific checks, see the Claude watermark detector. For Google watermarking, see the SynthID text detector and remover. For a broader Claude style discussion, see the Claude AI detector.

How the check works, step by step

The planned free detector is designed to make the uncertainty visible rather than hide it behind a single confident label. You paste the text, the system checks for several kinds of mark, and the result explains what was found.

1. Paste the text you want to check. Markdown and HTML formatting are preserved where possible, so you can inspect the text without stripping the structure that matters to you.
2. Run the check. Unmarker looks for invisible characters, formatting artefacts and statistical watermark signals where a relevant vendor watermark may apply.
3. Read the result. The output shows the likely signal, the level of uncertainty and the reason the result may be weaker for short, edited or translated text.

Unmarker is also building a rewriting workflow for Claude-marked text. That workflow rewrites with a model that carries no watermark, checks every sentence for meaning, including entities, numbers and quotes, then shows a before and after watermark score and a diff. If you want the broader cleaning workflow, see the AI text cleaner.

The free detector is planned as part of the pre-launch product. Early access is available through the waitlist, and the contact address is contact@alexandrefuchs.fr. The founder is Alexandre Fuchs, STRATINET SARL, RCS Cannes 944 837 871.

Why short and edited texts stay uncertain

Statistical watermarks need enough text to measure a pattern. A short paragraph may not contain enough choices for a stable reading. This is why a result should be treated as evidence, not as a final judgement.

Editing also matters. Anthropic says the Claude watermark may persist through light editing, but heavy paraphrase, translation or very short passages weaken it. That does not mean the text becomes clean in every sense. It means the measurable signal becomes less reliable.

A detected Claude mark means Claude touched the text, not that Claude wrote it from start to finish. Human text that was proofread or translated through Claude can be flagged. This is one reason false positives must be handled carefully.

The same caution applies to negative results. A low or uncertain score does not prove that no model was involved. It only says that the detector did not find enough of the measured watermark signal in the text provided.

AI watermark checker vs AI detector

An AI watermark checker answers a narrow question: does this text carry a vendor's key-based signal, or another technical mark that can be measured in the text. It is not trying to decide whether the style feels machine-written.

A style-based AI detector asks a different question: does this read like a model wrote it. Those systems have documented false positives. They may react to simple prose, non-native writing, edited text or formulaic professional language.

Unmarker does not claim to pass GPTZero, Turnitin, Originality or any other stylometric detector. It measures watermark signals and technical artefacts. That distinction matters if you are auditing content, checking publication risk or cleaning formatting debris before reuse.

If you need a removal workflow for watermark signals rather than only a check, the related AI text watermark remover page explains the planned rewriting approach. The detector page stays focused on inspection and uncertainty.

Planned pricing

The detector is planned to include a free tier. The current plan is Free with 3 texts per day, up to 1,500 words. Pro is planned at $9 per month, with $5 per month locked for the first 500 founders.

An API is planned at $0.50 per 10k words. These prices may change before launch. The product is pre-launch, so the safest next step is to join the waitlist if you want early access.

The planned pricing does not change what the tool claims to do. It checks text watermark signals and related technical marks. It does not promise an outcome with style-based AI detectors, and it should not be used as a substitute for context, records or editorial judgement.

Questions people ask