This week,Anthropic has confirmed that Claude models will now enforce a worldwide roll-out of invisible, machine-readable watermarks embedded into its AI-generated content. Under the EU AI Act’s Transparency Code, which took effect on 2 August, AI companies must introduce ways of identifying synthetic content.
Every Claude model launched on or after 2 August will now feature the watermark once generated, inclusive of text and imagery. Older models will undergo a transition period. Such marks will be invisible to humans, though fully detectable by automated systems. The marking happens at model level, meaning it travels regardless of whether you’re in the API, chat app, Claude Code, Claude Cowork, Claude Tag, or otherwise accessing it through cloud services like AWS, Microsoft Foundry or Google Cloud.
The EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content is a response to global concern that AI-generated content is becoming increasingly difficult to decipher from the human stuff. As AI slop continues to dominate internet corners, regulators continue to put pressure on the broader industry to watermark AI content.
Anthropic joins a line-up of other conglomerates also committed to complying with the EU’s transparency code, including Google, Meta, Microsoft and OpenAI. Signing the Code of Practice on Transparency of AI-generated Content is currently voluntary.When Claude will generate a supported image, Anthropic will attach digitally signed provenance metadata, whilst its text outputs will now hold what the company refers to as an “imperceptible watermark”. The signed C2PA metadata attached to image files are more fragile mechanisms, which can be stripped by screenshot or a conversion of the file format. But text watermarks may be more difficult to shake.Let’s not forget, it’s an interesting time for any realm of text on the internet, with the line between human and machine writing and editing continuing to blur exponentially. As such, the new watermarking system itself may carry varying degrees of reliability when it comes to identifying what’s fully AI generated, and what’s not. Any piece of text generated by supported Claude models will carry the watermark even as the user copies and pastes it, and the mark may even hold up through some degree of editing.
Anthropic is pretty clear about this limitation, noting that “Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.”
It also warns that the absence of a mark does not necessarily prove that AI wasn’t involved. This could be due to heavy editing, which will interfere with the signal, or in instances of short text which might not have enough material to be detected. Claude states that detection mechanisms are coming “in forthcoming technical documentation”.
At the time of writing, Anthropic has not yet confirmed exactly how the watermark system works. After all, we’re heading towards an internet where AI isn’t necessarily something you can see. The ambition to make it traceable without making it visible was, perhaps, a matter of time.
Every Claude model launched on or after 2 August will now feature the watermark once generated, inclusive of text and imagery. Older models will undergo a transition period. Such marks will be invisible to humans, though fully detectable by automated systems. The marking happens at model level, meaning it travels regardless of whether you’re in the API, chat app, Claude Code, Claude Cowork, Claude Tag, or otherwise accessing it through cloud services like AWS, Microsoft Foundry or Google Cloud.
The EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content is a response to global concern that AI-generated content is becoming increasingly difficult to decipher from the human stuff. As AI slop continues to dominate internet corners, regulators continue to put pressure on the broader industry to watermark AI content.
Anthropic joins a line-up of other conglomerates also committed to complying with the EU’s transparency code, including Google, Meta, Microsoft and OpenAI. Signing the Code of Practice on Transparency of AI-generated Content is currently voluntary.When Claude will generate a supported image, Anthropic will attach digitally signed provenance metadata, whilst its text outputs will now hold what the company refers to as an “imperceptible watermark”. The signed C2PA metadata attached to image files are more fragile mechanisms, which can be stripped by screenshot or a conversion of the file format. But text watermarks may be more difficult to shake.Let’s not forget, it’s an interesting time for any realm of text on the internet, with the line between human and machine writing and editing continuing to blur exponentially. As such, the new watermarking system itself may carry varying degrees of reliability when it comes to identifying what’s fully AI generated, and what’s not. Any piece of text generated by supported Claude models will carry the watermark even as the user copies and pastes it, and the mark may even hold up through some degree of editing.
Anthropic is pretty clear about this limitation, noting that “Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.”
It also warns that the absence of a mark does not necessarily prove that AI wasn’t involved. This could be due to heavy editing, which will interfere with the signal, or in instances of short text which might not have enough material to be detected. Claude states that detection mechanisms are coming “in forthcoming technical documentation”.
At the time of writing, Anthropic has not yet confirmed exactly how the watermark system works. After all, we’re heading towards an internet where AI isn’t necessarily something you can see. The ambition to make it traceable without making it visible was, perhaps, a matter of time.