Claude Adds an Invisible Watermark to Copy-Pasted AI Text

The change does not stop users from copying Claude's responses or from placing a visible label on them. Instead, Anthropic says the watermark is embedded directly into the generated text without changing its meaning, quality, or readability.

The change is tied to the European Union's new AI transparency requirements. Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. Article 50 sets out requirements to ensure AI-generated or manipulated content is identifiable in a machine-readable format.

Anthropic says Claude models launched on or after August 2, 2026, support its marking system from launch. Older models are expected to gain marking capabilities during a transition period, but previously generated text will not be watermarked. Despite the European rules driving the change, Anthropic is applying the system worldwide.

Unlike a watermark stamped across an image, Claude's text watermark cannot be seen during normal reading. Anthropic describes it as an "imperceptible watermark" woven directly into generated text.

Because the mark exists within the text itself, it can travel with a response when someone copies and pastes it into another location. Anthropic says it may also survive some editing.

Anthropic says marking occurs at the model level, meaning supported output can carry the watermark through Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag. It also applies to supported Claude models accessed through AWS, Google Cloud, and Microsoft Foundry.

Once Anthropic releases its detection tools, schools, publishers, employers, and online platforms could gain another method to identify potential Claude involvement. But the watermark's limitations mean it cannot provide a simple answer to who actually wrote something.

For more articles like this, check out our Tech News page!