Copy a paragraph from Claude into Word and it looks like ordinary text. There’s no Claude logo, warning label, or obvious marker attached to it. But the text may still contain a watermark.
Anthropic has started adding invisible watermarks to text generated by supported Claude models. It’s also adding provenance information to certain files, including images. The change is tied to new European transparency rules for AI-generated content that took effect on August 2, 2026.
The part that may surprise people is that the watermark isn’t something you can see in the document.
Claude generates text by choosing one word or token at a time. In many situations, several options would work. Take a sentence like:
“Her presentation was clear and…”
Claude could continue with “concise,” “well organized,” or “easy to follow.” Each option fits the sentence.
Anthropic’s watermarking system takes advantage of those choices. Across a longer passage, Claude can favor certain word options in a way that creates a statistical pattern. To the reader, the writing still looks normal. A detector with the right key can look for that pattern later.
There are no special characters hidden in the document. Copying the text into Word, PowerPoint, or an email doesn’t automatically remove the signal.
Anthropic says some editing may leave the watermark intact, while heavier rewriting can make it harder to detect.
The system is based on SynthID-Text, a technology developed by Google DeepMind. Google already uses SynthID across several kinds of AI-generated content, including text produced with Gemini.
Not yet.
Anthropic says it’s working on a watermark detection API and plans to release more technical information for third parties. As of August 28, 2026, there’s no public Anthropic tool where someone can paste text and check it themselves.
When detection becomes available, it won’t work like a simple “AI” or “human” label. The system is meant to estimate whether Claude’s watermark is present in the passage.
Longer text is easier to evaluate because the detector has more word choices to analyze. Short passages are more difficult. The same is true for code, highly factual writing, and light editing, where Claude has less freedom to choose between different ways of saying the same thing.
A failed detection also wouldn’t prove that a person wrote the text. The passage might have come from another AI tool, an older Claude model, a short response, or something that was heavily rewritten.
The reverse is also true. Detecting Claude’s watermark only shows that Claude was involved in producing the text. Someone could have written the original material themselves and then used Claude to edit, translate, or revise it.
Anthropic is using a separate approach for supported files such as PNG, JPG, and SVG images.
These files can include signed provenance information using the C2PA standard. That information can record details about where the file came from and whether Claude processed it. This differs from the text watermark. The text signal is built into patterns of word choice, while C2PA information travels with the file itself.
The timing is linked to the European Union’s AI Act. Article 50 introduced transparency requirements for certain AI-generated and AI-manipulated content. Those requirements became applicable on August 2, 2026.
The EU also created a Code of Practice to help AI companies comply with those rules, and Anthropic is one of the organizations that signed it.
There is one important detail for users outside Europe: The regulation comes from Europe, but Claude’s watermark is being applied globally. Anthropic says it’s not currently limiting the system to European users because it doesn’t have a durable way to apply the watermark by region.
That means someone using a supported Claude model in the United States, Canada, Australia, or elsewhere may still receive watermarked text.
Claude isn’t the only AI product moving in this direction.
Google already uses SynthID for several types of AI-generated content and has applied SynthID-Text to text generated through Gemini. Anthropic’s system is based on Google’s published approach.
Other major AI companies, including OpenAI, Microsoft, Meta, Mistral, and Cohere, have also signed the relevant part of the EU’s transparency Code of Practice.
OpenAI already uses provenance and watermarking technologies in some types of generated media. Its public information doesn’t currently say that normal ChatGPT text uses the same kind of text watermark Claude is introducing.
Microsoft is also using provenance and watermarking in parts of Microsoft 365, particularly for generated media.
The broader trend is clear, even if the technology varies from company to company. More AI-generated content is likely to carry some form of machine-readable provenance or marking.
For users, that creates a new way to think about AI detection. Most AI detectors today try to judge whether a piece of writing looks AI-generated based on style and probability. Claude’s watermark works differently because the signal is placed into the text during generation.
© 2026 SVA Consulting