New Delhi: Anthropic announced a watermarking system for content generated through Claude, allowing machine systems to identify text, images and files produced with its AI models.
The change covered Claude models launched on or after August 2 and applied across Claude, its API, Claude Code, Claude Cowork and other supported surfaces. Anthropic planned to use the system globally, not only within the European Union.
The move came as the company worked to comply with the EU’s Code of Practice on Transparency of AI-generated Content, which required AI providers to make synthetic content easier to identify.
For text, Claude would insert signals that remained invisible to readers but could be detected by computer systems.
These markers could remain embedded when users copied and pasted the text between applications or platforms.
“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 said.
The company said it would provide more information later on how the watermarks could be detected.
The system did not rely on a visible label attached to a paragraph. Instead, the watermark formed part of the generated text and was designed to survive ordinary copying and pasting.
Claude-generated image files, including formats such as PNG, JPG and SVG, would carry digitally signed provenance metadata.
The metadata could include information about the origin of the file, how it was created and whether it was modified later.
A cryptographic signature would also help indicate if somebody tampered with the original metadata.
The approach was intended to make it easier for platforms and other systems to establish whether a file originated from an AI tool.
Anthropic also cautioned that the system would not provide a definitive test of whether Claude played a role in producing a piece of content.
A watermark could appear in text that started with human-written material but was later rewritten, summarised, translated or proofread using Claude.
That created an important distinction between AI-generated work and AI-assisted work.
For example, a document could contain only a small amount of Claude-generated text but still carry a detectable marker. On the other hand, extensive editing, paraphrasing, or combining AI output with other writing could weaken or remove the watermark.
The same limitation applies to images. If a Claude-generated image was captured through a screenshot and saved as a new file, the original provenance metadata could disappear.
Anthropic said these limitations meant that the presence of a watermark would not always prove that Claude created an entire piece of content. Its absence would also not necessarily prove that Claude was never involved.
Anthropic’s move came as technology and content platforms experimented with different ways of disclosing AI-generated material.
The European Union’s transparency framework pushed AI companies towards clearer identification of synthetic text, images and other media, including deepfakes.
The challenge remained particularly difficult for text because users could easily edit, translate or combine machine-generated material with their own writing.
For Anthropic, the watermarking rollout marked an attempt to create a technical signal around AI use while acknowledging that the system could not serve as a foolproof AI detector.