Claude will hide invisible watermarks in generated text
Key Points
- Claude will begin embedding invisible watermarks into generated text to flag to users that it has been AI-generated.
- After signing the EU's Code of Practice on Transparency of AI-Generated Content, Anthropic has begun rolling out machine-readable marking to supported content generated by its new models.
- Users will not be able to see the hidden watermarks in text and image files, but tools will be able to detect that content has been wholly or partly AI-generated.
- There are still ways to circumvent this marking mechanism, but they will make it more difficult for those seeking to pass off AI-generated output as their own work.
Anthropic has announced that its Claude models will begin inserting invisible watermarks into AI-generated text, making it easier for people to check if what they are reading has been AI-generated.
The company has signed the EU’s Code of Practice on Transparency of AI-Generated Content, which obliges providers of generative AI models or systems to embed watermarks or other indicators in generated content that allows output to be detected as generated by AI.
In a post on its Claude support page, Anthropic said that Claude models launched in the EU on or after 2 August 2026 will support machine-readable marking at launch.
This means that text generated by these models will carry invisible, embedded watermarks and other files or content will carry metadata that describes the content as AI-generated.
Users will not be able to see the watermark directly; it will be woven imperceptibly into the text and will not change the quality or meaning of Claude’s output.
This watermark will travel with the text when it is copied and pasted, and it will be present no matter which platform or surface is used to generate content using Claude.
Additionally, supported image files will also attach signed provenance metadata, which will signal that a file has been processed by Claude.
Detection tools are on the way
Anthropic said it is currently working on ways to allow users and third parties to detect this embedded watermark and provenance metadata in other files.
This mechanism will check whether a piece of text carries the invisible watermark and, if it does, it will flag that the content the user is seeing may have been processed by Claude.
Once Anthropic releases documentation on its detection mechanisms for these invisible watermarks and image metadata, users can expect new extensions or online tools that should make detecting and flagging Claude-generated content much easier.
Anthropic notes that if a Claude mark is detected, it does not mean the content has been wholly authored by the LLM. The original prompter may have used Claude to proofread, convert, or translate their work.
Additionally, there will be ways to circumvent this watermarking. If text has been generated by a model without this capability, it will not include the invisible watermark.
Text passages that are too short, or which have been heavily edited and paraphrased, may also no longer carry a detectable mark.
File metadata, such as that used to identify AI-generated images, can also be stripped using various tools and workarounds such as screenshots.
While this solution is not a perfect one for addressing the deluge of AI-generated misinformation and content posing as original work online, it is a promising signal for those frustrated by the proliferation of ‘AI slop’ online.
