Anthropic has announced plans to mark content generated or processed by Claude as part of its commitments under the EU AI Act and the Code of Practice on Transparency of AI-Generated Content.
For new Claude models launched in the EU on or after August 2, 2026, Anthropic says marking will be supported from launch.
The approach relies on two complementary mechanisms: embedded watermarks in text and signed provenance metadata for files.
An Invisible Watermark Embedded in Text
This is perhaps the most interesting part of the announcement.
When a supported Claude model generates text, Anthropic says it will embed an imperceptible watermark directly into the text.
This is not simply a visible “Generated by AI” label. According to Anthropic, the watermark travels with the text when it is copied and pasted and may remain detectable after some editing.
How is that possible?
Anthropic has not yet published the technical details of its implementation and says further technical documentation will follow.
One possible approach to this type of watermarking is to introduce a statistical signal during text generation. A model can favor certain token choices or formulations among several possible alternatives, creating a statistical pattern that can later be detected across a sufficiently large sample of text.
Such an approach would explain why a watermark could survive limited editing while becoming increasingly difficult to detect after substantial rewriting.
Copying the Text May Not Remove the Mark
This has an important consequence.
A simple copy and paste preserves the text and may therefore preserve the watermark.
Limited modifications may also leave enough of the signal intact for detection.
However, Anthropic explicitly acknowledges several limitations. Heavy editing, paraphrasing, translation or combining the content with other text may make the watermark undetectable.
Very short passages may also contain too little information to provide a reliable signal.
A text watermark should therefore be considered a provenance signal rather than absolute proof of authorship.
C2PA Provenance Metadata for Files
For supported files, Anthropic is taking a different approach.
Claude will attach signed provenance metadata to supported file types such as SVG, PNG and JPEG.
This metadata follows C2PA (Coalition for Content Provenance and Authenticity), an open industry standard designed to provide verifiable information about the origin and history of digital content.
When signed provenance metadata is present, it can indicate that a file was processed by Claude and help detect whether it has subsequently been altered.
This approach is particularly interesting for images and other digital assets because it moves the discussion beyond simple AI detection toward broader content traceability and provenance.
It also has limitations. Metadata can disappear when files are converted, re-saved, screenshotted or processed by systems that do not preserve it.
“Processed by Claude” Does Not Mean “Created by Claude”
This is probably the most important distinction in Anthropic’s announcement.
Detecting a Claude mark does not necessarily mean that Claude originally created the content.
Consider a simple example.
A person writes an entire article and then asks Claude:
“Correct the spelling and improve the style slightly.”
The resulting text may carry a Claude mark even though the ideas and original content came from a human author.
The same applies to translation, summarization, proofreading or file conversion.
It would therefore be incorrect to interpret a detected mark automatically as:
“This content was written by AI.”
A more accurate conclusion would be:
“This content may have been processed by Claude.”
That distinction matters.
No Watermark Does Not Mean “Human”
The opposite is equally important.
The absence of a detectable watermark does not prove that content was created by a human.
Anthropic identifies several situations where Claude-generated or Claude-processed content may no longer carry a detectable mark:
it was generated by a model that did not yet support marking;
the text was heavily edited or paraphrased;
it was translated;
it was combined with other content;
the passage is too short;
file metadata was removed;
the platform, feature or file format did not support a particular marking mechanism.
This highlights a fundamental limitation of AI-content detection.
These mechanisms can provide signals, but those signals should not be treated as definitive proof.
From AI Detection to Content Provenance
This development is particularly interesting because it may gradually change the question we ask about digital content.
Today, the common question is:
“Was this text or image created by AI?”
But that distinction is becoming increasingly difficult to establish — and increasingly less meaningful.
Content can be:
written by a human and corrected by AI;
generated by AI and extensively rewritten by a human;
translated by AI;
summarized by AI;
assembled from multiple human and AI-generated sources;
or processed successively by several AI systems.
A simple Human vs. AI classification quickly reaches its limits.
We are therefore gradually moving from:
“Is this AI?”
to:
“What is the provenance of this content?”
That is a much more meaningful question.
A New Dimension of AI Governance
For organizations, this development goes far beyond identifying AI-written articles on the Internet.
As generative AI becomes integrated into business processes, organizations will increasingly need to understand — and in some cases demonstrate:
which AI systems were used;
which content was generated or transformed;
where human intervention occurred;
what modifications were made;
what provenance evidence is available.
In other words, traceability is becoming an important component of AI governance.
This is a familiar principle in cybersecurity and GRC. Organizations cannot rely solely on statements that a process is controlled. They increasingly need evidence showing how processes operate and, where necessary, the ability to trace what happened.
Technologies such as watermarking and C2PA can contribute to that traceability, although they cannot provide the entire solution on their own.
An Important Step, but Not a Perfect Solution
Anthropic’s initiative represents an interesting step toward greater transparency around AI-generated content.
But its limitations are just as important as the technology itself.
A watermark can become undetectable.
Metadata can be removed.
Marked content may originally have been created by a human.
And unmarked content may still have been generated by AI.
The real objective may therefore not be to build a universal detector capable of answering “human or AI?” with certainty.
Instead, the challenge is to develop an ecosystem of provenance, traceability and governance for digital content.
As the EU AI Act progressively takes effect, this topic is likely to become increasingly important — both for AI model providers and for the organizations using these systems.
AI content watermarking: a real step forward for transparency, or a measure that will ultimately be easy to circumvent?
Top comments (0)