Anthropic Adds Hidden Watermarks to Claude AI Text to Help Identify AI-Generated Content

Anthropic on Tuesday announced that it is adding invisible watermarks to text generated by its artificial intelligence (AI) models. According to the company, the content will have a hidden signal that remains even when it is copied and pasted elsewhere. The move builds upon Anthropic's decision to sign the European Union's AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content. Anthropic says Claude models launched from August 2, 2026, will support the marking system from day one, and the feature is rolling out globally.

Invisible Watermarks in Claude

Anthropic updated its support page to reveal a list of commitments it is putting into practice, in line with the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, which it has signed as a provider of both generative AI models and generative AI systems.

As per the company, the watermark will be embedded within text generated by supporting Claude models. It has been designed to travel with the text when copied and pasted elsewhere, and can also survive some level of editing. While how the detection system will work is still being finalised, Anthropic eventually aims to provide users and third parties with tools to detect the embedded watermarks.

Anthropic, however, emphasised that it will be imperceptible to users. Further, the addition of a watermark will not affect the meaning, quality or readability of the response. The process is claimed to be carried out at the model level and is not tied to a particular Claude product or interface.

The company also claims that the watermark will not be a definitive AI detector, and should not be treated as such. “Detecting a Claude mark tells you that the content may have been processed by Claude. It does not, on its own, confirm the full provenance of the content,” the support page reads.

Citing an example, Anthropic said a user could write an article themselves and then use Claude to either proofread, translate, summarise, or change its formatting. Even then, there is a chance that the result could carry a Claude watermark, despite the initial idea and writing originating from the user.

Meanwhile, the opposite is also likely; a missing watermark may not prove that content was not generated or processed by AI. Heavy editing, paraphrasing, translation, combining Claude output with other writing, or using a very short passage could make the mark undetectable, the company said.

The new watermarking system is said to cover supported Claude models across Claude, Claude Platform (API), Claude Code, Claude Cowork and Claude Tag. Such watermarks will also be applied when supported models are accessed through cloud platforms including AWS, Google Cloud and Microsoft Foundry, where supported.

Anthropic will use digitally signed provenance metadata for supported file types like SVG, PNG, and JPG, based on the Coalition for Content Provenance and Authenticity (C2PA) standard. This metadata will indicate if a file has been processed by Claude, and whether the provenance information has been tampered with.

The watermarking system is being introduced in markets where supported Claude models are available.