Anthropic is preparing to make AI-generated content from Claude easier to identify, with plans to embed invisible markers into text and add provenance information to supported images. The changes are intended to meet new transparency requirements under the European Union’s AI Act, but Anthropic says the marking system will be deployed globally rather than restricted to Europe.
For images, Anthropic plans to use C2PA, the open provenance standard designed to attach information about a digital file’s origins and editing history. The technology has increasingly become an industry-wide approach to AI content transparency, with companies including Adobe, Google and OpenAI supporting C2PA-based credentials in various products.
Text presents a more complicated problem. Anthropic says Claude-generated writing will contain an imperceptible watermark embedded directly into the output. According to the company, the watermark should not affect the meaning, quality or readability of the generated text.
Because the marker is part of the text itself, Anthropic says it can remain attached when someone copies and pastes Claude’s output and may survive at least some subsequent editing. The watermarking will operate at the model level rather than being limited to Anthropic’s consumer chatbot.
That means supported Claude models should produce marked text when accessed through Claude products as well as third-party infrastructure, including Amazon Web Services, Google Cloud and Microsoft Foundry. Anthropic says the approach will cover services such as its API, Claude Code and Claude Cowork.
The company has not yet provided detailed technical information explaining how the text watermark works or how resilient it will be against extensive rewriting. Anthropic says it intends to publish documentation that will allow users and third parties to detect its watermarks and provenance information.
The rollout also won’t happen across every Claude model immediately. New models are expected to include content marking from launch, while Anthropic is working on adding the technology to existing models. The timing follows AI transparency obligations under the EU AI Act that took effect on August 2, with existing products receiving a four-month compliance period.
The broader challenge is whether invisible AI labeling can remain reliable once content leaves the system that created it. C2PA credentials, for example, can disappear when files are processed by services that don’t preserve the metadata. Text watermarking faces its own potential problems if content is heavily edited, translated or rewritten.
That makes Anthropic’s move more useful as a provenance signal than definitive proof of authorship. The absence of a detectable watermark cannot establish that something was written by a human, just as finding one would depend on detection tools being accurate and widely available.
Still, embedding provenance at the model level could make AI identification more consistent than relying on platforms to voluntarily label material after publication. As regulators push technology companies toward greater transparency, the bigger test will be whether these systems can survive the messy reality of copying, editing and sharing content across the internet.
