Last Updated: 13 August 2026
Anthropic is adding machine-readable marks to content made by supported Claude models, as part of its work to meet EU AI Act transparency rules. The change brings an invisible watermark for generated text and signed provenance metadata for some files. Newer Claude models support this marking from the start, while Anthropic is still rolling support out to older models.
This is not just a European rule. Anthropic says supported Claude models will carry marks everywhere, across Claude and its API products. The same text watermarking can also show up through supported deployments on major cloud platforms. That makes this change matter for everyday users, developers, and businesses using Claude outside Europe too.
The key thing to understand is what the mark actually proves. This detected mark means the content may have passed through Claude at some point. It does not prove that AI wrote the original text from beginning to end. Text written by a person can still carry this signal after Claude works on it, which makes the watermark more of a clue about where content came from than a simple test of who wrote it.
A Claude watermark is not proof that the AI wrote a piece of text. Anthropic calls it a provenance signal instead. A positive result can show that Claude worked on the content, but the material might have existed before it got involved. The original ideas or wording could still come from a person or another source entirely.
That difference is important because content often passes through several stages. Claude might work on something that was already written, and the final version can still carry its mark. The text might also get changed later, or mixed with other material. A watermark cannot rebuild that whole history, so nobody should treat it as a full record of who wrote what.
“Knowing a watermark is present tells you content passed through an LLM at some point. It doesn't tell you whether that content is accurate, on-brand, or credible enough to publish under your company's name"
Kayce Danna, VP of Product at Markup AI
The safest way to read it is narrower than it sounds. A detected Claude mark can support the claim that Claude touched the content in some way. It cannot support the stronger claim that Claude generated every sentence or came up with the original ideas. This mix-up is easy to make, because the phrase "AI watermark" sounds more certain than what Anthropic actually says it means.
A Claude mark points to possible involvement. It does not settle who wrote the original material, how much Claude added, or what happened to the text later on.
Anthropic says supported Claude models build an invisible watermark directly into generated text, right at the model level. Readers should not notice anything different in normal use. The mark should not change what the response means or how easy it is to read. Because the signal lives inside the text itself, it tends to stay with the content even after someone copies and pastes it elsewhere.
Claude uses a different system for some files it generates. Supported PNG, JPG, and SVG files can get signed C2PA provenance metadata attached to them. That metadata can help show a file passed through Claude, and whether it was changed afterward. This should not be confused with the invisible text watermark, since Anthropic describes it as a separate signal working at the model level.
There is still a big technical gap here. Anthropic explained what the text watermark is supposed to do, but it has not shared the method behind it in detail. That means readers should not assume Claude works the same way as Google's SynthID. Anthropic has not named its text-watermarking system, nor explained exactly how it creates the signal.
Claude uses an invisible signal inside text, plus signed metadata for some files. Both aim at similar provenance goals, but Anthropic has not shown that they work through the same technical method.
Text written by a human can still end up carrying a watermark. Anthropic gives proofreading, translation, and summarizing as examples of work Claude might do on content that came from somewhere else. Someone could hand over their own draft and get back processed text that still carries a Claude signal. The mark would show that AI was involved, but it would not prove Claude came up with the original ideas.
This lines up with an important detail in the EU rules too. The European Commission says the marking duty placed on providers does not apply when an AI system only performs a basic editing role. That legal point stands apart from how Anthropic designed its own product. It does not tell us whether every editing task inside Claude will end up marked or not.
This distinction could be highly relevant a lot for schools, employers, and publishers if Claude detection becomes common. A mark might show that AI touched a document without proving AI wrote it. These groups may need to separate AI assistance from AI authorship before using a detected signal to judge how a piece of writing was really made.
Copying and pasting should not remove Claude's text watermark, since the signal sits inside the text itself. The company clarified that the mark can survive some editing. Even so, it has not shared a clear line showing how many changes a piece of text can take before detection stops being reliable, or the signal fades away completely.
Anthropic already points to a few situations where detection can fail. Heavy rewriting can weaken the signal a lot. Paraphrasing or translating the text can also make it harder to find. Very short passages might not contain enough material for a dependable result. All this means a missing watermark cannot be treated as proof that Claude, or any other AI tool, played no part in the text.
Files carry a different kind of weakness. Signed provenance metadata can vanish when a file gets converted or saved again. Taking a screenshot can wipe that information out too. These file-related limits should be kept separate from text editing, since the two systems work in different ways. Anthropic still has not explained how well the text watermark holds up after major changes.
“We are now entering a period where content creators and AI models will be locked in an arm's race as the models seek to distinguish their content and content creators seek to remove all traces of AI from their writing.”
Kadan Stadelmann, CTO and Co-Founder of Compance.AI
A positive mark does not prove full AI authorship, and a missing mark does not prove a human wrote everything either. Both results need context before anyone uses them to judge a document.
Newer models support marking right from launch, and the company is still adding this support to older ones. Under the EU rules, systems that were already on the market before the cutoff get a short grace period before the specific marking and detection duty kicks in. This gives older qualifying systems some breathing room to catch up.
The rollout goes well beyond Europe. Supported models will carry marks wherever the assistant is offered. This covers the main chat product, along with its developer platform and API. The coding tool is part of the plan too, alongside the team-collaboration and tagging products. That means both everyday users and organizations building on top of these models will see this change as the rollout continues.
Text watermarking can also show up through supported deployments on major cloud providers like AWS and Google Cloud. Microsoft Foundry is included as well. Signed file provenance may look different across platforms, since not every one of them supports every type of marking. The company also tells developers to check their own compliance duties, since the built-in marks alone do not settle every question for products built on top of these models.
Claude is not the only major AI product using provenance signals for text. Google already uses SynthID-Text for content generated through the Gemini app and web experience. Google has also published research explaining how its method works. Anthropic has not done the same yet, so the two systems should not be treated as if they work the same way.
OpenAI sits at a different stage right now. Its current provenance documentation lists supported signals for images and audio, but not for text. OpenAI has said it wants to expand these signals to cover every format, including text eventually. That makes the comparison more useful than simply saying ChatGPT has no watermark at all, since the company has already stated where it wants to go next.
The main takeaway here is that there is no single shared standard for text watermarking across Claude, Gemini, and ChatGPT right now. Google has a documented system for text. Anthropic has announced a model-level watermark without publishing its technical method. OpenAI currently documents other formats while planning text support down the line.
We already know what its watermark is supposed to signal, but several practical details are still missing. The company says it is working on letting users and outside groups detect Claude's embedded marks. It has not yet published full technical documentation for that detection system, or explained exactly how people will get access to it.
The biggest open question is how tough the signal really is. We still don't know how much editing a marked passage can take before detection becomes unreliable. They didn't provied yet a Claude-specific numbers for false positives or false negatives. There is no public minimum text length for dependable detection either, though Anthropic warns that very short passages may not carry enough signal.
These gaps do not make the watermark useless, but they do limit what anyone can safely conclude from it right now. Anthropic already admits several cases where detection can fail, and its detector is not fully documented yet. Until more technical details come out, the safest way to describe this is simple: Claude's watermark is a provenance signal with limits, not a final answer about who wrote a text.
No. Anthropic says watermarking applies to supported Claude models, and support for older models is still being added. Newer models support marking from launch. The system is also meant to work across Claude products and supported cloud deployments around the world. Still, not every older model is covered yet. That is why it is more accurate to say supported models watermark generated text, rather than claiming every response already carries a mark you could detect today.
No. A detected watermark can show that Claude worked on the content, but it does not prove AI wrote all of it from scratch. Anthropic says the material may have started somewhere else before Claude touched it. That includes human-written content used for tasks like proofreading or translation. The text may also change later on. A positive result is really a provenance signal about Claude's involvement, not a full record of who created every sentence or idea.
It can make the watermark harder or impossible to find, but Anthropic has not published a clear threshold for removal. Copying and pasting should keep the signal intact. Some editing can also leave it detectable. Heavy rewriting can weaken it, while paraphrasing or translating the text may cause detection to fail. Very short passages can be unreliable too. Since Anthropic has not said exactly how many changes it takes, any claim about a specific amount of editing removing the watermark should be treated with caution.
OpenAI does not currently list text among the supported content types on its public provenance page. Its documented signals cover supported images and audio only. That said, OpenAI has said it wants to expand provenance signals to every format, including text. That is a more specific answer than simply saying ChatGPT will never watermark text. Claude is already rolling out an embedded text watermark for supported models, while OpenAI has publicly stated that text provenance is part of where it wants to go next.
There is no confirmed evidence that Claude uses Google's SynthID system. Google already uses SynthID-Text for content generated through the Gemini app and web experience, and it has published research explaining the method behind it. Anthropic has described its own watermark as an invisible, model-level signal, but it has not named the technique behind it. Both systems deal with text provenance, but they should not be treated as technically identical unless Anthropic confirms otherwise.
