Anthropic adds invisible watermarks to AI text to boost transparency

AI company Anthropic has announced that its Claude models can now embed invisible watermarks in AI-generated text, creating a hidden marker that remains even after the text is copied or edited. The move aims to help users and platforms identify whether a piece of content was produced by AI.

According to a support page updated by the company, supported Claude models can embed a machine-readable watermark directly into text without altering its readability. The marker is imperceptible to readers but can be detected by tools, and it survives copy-paste and some editing.

Anthropic said the feature is part of its commitments under the European Union's AI Act, specifically the Code of Practice on Transparency of AI-Generated Content. Claude models launched in the EU on or after August 2, 2026, will support the machine-readable marking from launch. The company added that the marking will apply globally to all supported models.

The watermark system covers Claude's API as well as its products, including Claude, Claude Code, Claude Cowork, and Claude Tag. In addition to text, Claude can also attach digitally signed provenance metadata to image files such as PNG, JPG, and SVG, using the C2PA standard. This metadata can indicate that the file was processed by Claude and help detect tampering.

However, Anthropic emphasised that a detected watermark is not conclusive proof of content origin. For instance, someone may have written original material and used Claude for proofreading, translation, summarisation, or conversion, in which case the resulting content could still carry a Claude mark.

Similarly, content can change after Claude processes it—through editing, excerpting, or mixing with other material. The absence of a watermark does not necessarily mean AI was not involved. According to the company, a Claude-generated passage may not carry a detectable mark if it came from an older model, was heavily edited or paraphrased, is too short for a reliable signal, or has been combined with other writing.

For files, provenance metadata can disappear when images are converted to different formats, re-saved, or screenshotted. Anthropic also said it is still developing the tools that will allow users and third parties to detect the watermarks and provenance metadata. More technical details on how detection works are expected to be published later.

This initiative is not an AI lie detector. Instead, it functions as a hidden digital trail, giving users and platforms another signal to consider when assessing whether content may have been AI-generated. As the technology evolves, and with regulations like the EU AI Act pushing for transparency, watermarking is likely to become a more common feature across the industry.