AI watermarking, explained honestly

AI watermarking is a signal a model embeds in what it generates so that a detector can later recognise the output as machine-made. For text the signal lives in the model's word choices; for images and audio it can live in the pixels or the sound, or ride alongside the file as signed metadata. A mark records that a model touched the content. It does not record who asked for it, or whether a person wrote the words first.

Since 2 August 2026, new Claude models weave an invisible statistical watermark into the text they generate, worldwide — Anthropic lists Fable 5.1, Mythos 5.1 and Opus 5 today, and says every earlier model will be covered by 2 December 2026[1]. It is the same scheme Google uses: Anthropic said on 14 August 2026 that Claude's mark is a version of SynthID-Text, the method DeepMind published in Nature in 2024[10][3][4]. The trigger is the EU AI Act, which from that date requires generative AI output to be marked in a machine-readable format[7]. Since 1 September 2026 a detector for that mark exists, but it is in private preview and you have to qualify for it[1]. Most of what you will read about removing these marks describes a different thing entirely.

This site tracks who marks what, by which method, and whether anyone outside the company can verify it. Everything here links to a primary source. Where we have not been able to confirm something, the page says so rather than guessing.

Where should you start?

• Who watermarks what — the tracker table. Provider, modality, method, and whether a detector exists.
• What Anthropic actually shipped — precisely, including the limits Anthropic states itself.
• Scan your text — a real check for hidden characters, run entirely in your browser, honest about what it cannot see.
• The detector audit — we looked at the tools ranking for "Claude watermark checker". None of them detect Claude's watermark.
• Flagged for something you wrote? — the case nobody is covering, and it is the most likely way this affects you.

What do people get wrong about AI watermarks?

Three different things get called a watermark, and most of the confusion comes from treating them as one.

1. Are invisible characters the same as a statistical watermark?

Most "AI watermark remover" tools strip zero-width Unicode characters — U+200B and its relatives — from your text. Those characters are real, and some tools do insert them. But they are not what Claude uses. Claude's mark lives in the model's word choices, not in extra characters[1], and Anthropic now states it flatly: nothing is added to the text and there are no hidden characters[10]. Deleting invisible characters does nothing to it. Our scanner checks for the former and tells you plainly that it cannot check for the latter.

2. Does a watermark mean the AI wrote it?

This is Anthropic's own caveat, and it matters more than the announcement did. A detected mark signals that content may have been processed by Claude, and is not fully conclusive, because people use Claude to proofread, translate, summarise and convert files[1]. Run your own essay through Claude for grammar and it carries the mark. What to do about that.

3. Can you use the detector?

On 1 September 2026 Anthropic put its watermark detection API into private preview[10][1]. It is offered to organisations that EU law obliges to check — regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups — and to enterprises under the same obligation, by application through a form[1]. Anthropic says access will widen over time[12]. If you are an individual wondering about your own text, that is not you, and there is still nothing you can run. Any site claiming to check your text for a Claude watermark today is checking something else.

Files are the exception, and the gap between the two is now the thing to understand. Anthropic runs a free Claude Content Checker that reads the C2PA credential on a file in your browser[1][11]. It takes images, video and audio. It does not take text[11].

This is not an unavoidable state of affairs, which is the part worth noticing. OpenAI runs a public verification tool for the images and audio it generates, checking both Content Credentials and SynthID, and has offered verification through an API since 31 July 2026[9]. Google open-sourced its text watermarking scheme along with a working detector[4]. Marking content, and deciding who is allowed to check it, are separable decisions, and different providers have made them differently.

Why does this site exist?

The rules changed on 2 August 2026 and the coverage that followed was mostly recycled from press releases. Within days the search results filled with tools selling removal of a watermark they cannot detect. There was no page that simply said what is marked, by whom, and what you can verify yourself. So we wrote one.

Corrections are welcome and get made. See about, and the reusable dataset.

Last verified against primary sources: 2026-09-18