Claude’s invisible watermarks on AI-generated text spark users backlash

Anthropic has introduced invisible watermarks to AI texts generated by new Claude models and the users are not happy with this move.

The US-based tech company took this major step as a part of its commitment under the EU AI Act’s transparency rules. It means that new models launched on and after August 2, 2026 will carry this machine-readable marking at launch.

How this marking works

According to Anthropic the marking will apply globally across all Claude products including Claude Code, Claude Cowork, Claude Tag and API. It will also be applicable through cloud partners like Microsoft Foundry, Google Cloud and AWS. Speaking about the regions, it will be applied to those where Claude services are being offered.

Claude will use two methods to mark the content. The first method is based on text watermarks by embedding an “imperceptible watermark directly into the text itself.”

If the text is copied or pasted, it will remain there. “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from,” Anthropic wrote in blogpost.

The second method is “signed provenance metadata attached to files” for the supported files types.

How these markings can detected

According to Anthropic, it is planning to release guidance which enables users and third parties to detect these marks. Organisations that deploy Claude within their own products will independently assess their own compliance with the act.

Here’s the catch

Here are some limitations of the marking. For instance, the marking does not provide conclusive proof that the text is generated by Claude as most of the users use the model to edit, proofread, translate or summarize the text.

“The content may have changed after Claude processed it. Marked content may be modified, excerpted, or combined with other material after Claude processed it,” it wrote.

On the other hand, if a text does not carry a mark, it does not prove that text is entirely human-written. The signals may be missing if the users use the model based on pre-dates marking support and text is heavily edited, translated, or too short for a reliable signal.

Based on the “copy-paste” clause and limitations, the users are questioning the efficacy of this water-marking.

One user wrote on X, “Such an awful feature. You can already sense it, you want to update a small text by adding 1 paragraph and it rewrites the whole file. Stop the nonsense surveillance. The EU does not get to decide for the whole world.”

Another commented, “I hope they get some pushback on this. If I'm paying for your plan, I don't want invisible watermarks embedded in my content. If this is really the direction they choose, we should seriously consider open source alternatives that don't impose this kind of control.”

“I’m confused how this affects anyone other than people who straight up copy and paste from Claude,” the third one wrote.