Claude Is Hiding a Signature in New AI Text. It Won’t Prove Who Wrote It

This is not a visible label on the page

Anthropic has published a new explanation of how Claude marks AI-generated content. The company describes two separate systems.

For text, Claude will embed an imperceptible watermark directly into the way the passage is generated. It will not appear as a badge, a warning, a special character or a note at the bottom of the page. Anthropic says it should not alter the meaning, quality or readability of the response. Because the signal is part of the text rather than ordinary file metadata, it may travel with the passage when someone copies and pastes it elsewhere.

For supported files such as SVG, PNG and JPG, Claude will attach signed provenance metadata using the C2PA standard. Think of that as a tamper-evident label describing where a digital file has been processed. It is not the same as the hidden mark in text, and it has a familiar weakness: metadata can disappear when a file is converted, re-saved, screenshotted or uploaded to a service that removes it.

The important date is 2 August 2026

The timing comes from Europe. The European Commission says that the transparency obligations in Article 50 of the AI Act apply from 2 August 2026. The Commission’s Code of Practice on Transparency of AI-generated Content is intended to help providers meet requirements for marking and labelling synthetic content.

Anthropic says Claude models launched in the EU on or after that date will support marking from day one. It has also chosen to apply the system worldwide, wherever Claude is offered, rather than creating one output for European users and another for the rest of the world.

That does not mean that every Claude response already carries a mark. Models released before the cutoff are covered by a transition period, and Anthropic says support for them is still in progress. The company has not published a clear public list showing which existing models are currently marked.

The marking is applied at the model level. Anthropic therefore says it will cover supported output from Claude itself, the API, Claude Code, Claude Cowork and Claude Tag. Text marks are also expected when supported models are accessed through AWS, Google Cloud or Microsoft Foundry. File provenance may vary according to what a particular platform supports.

A signature is not a confession

There is a danger in describing this announcement as a perfect way to identify AI-written text. Anthropic does not describe it that way. The company says a detected mark is a signal that content may have been processed by Claude. It does not prove that Claude was the original author.

That distinction matters in ordinary life. Imagine that a student writes an essay, then asks Claude to correct the grammar. Or a doctor dictates a letter and uses an AI tool to translate it. Or a researcher supplies their own notes and asks Claude to turn them into a readable summary. The finished text may carry a Claude mark even though the ideas, facts and responsibility came from a person.

The opposite is also true. If no mark is detected, that does not establish human authorship. Anthropic lists several reasons why a mark might not be found: the passage may be too short, the text may have been heavily edited, translated or paraphrased, or it may combine Claude’s output with other writing. In a file, the provenance metadata may simply have been removed during normal handling.

In other words, the system may answer a narrow question — “Could Claude have processed this?” — but not the much larger question — “Who wrote this?” Treating the first answer as the second would create exactly the kind of unfair shortcut that current AI detectors have already encouraged.

Why some users are worried

The first public reaction was not uniformly enthusiastic. Some users welcomed a reliable signal that could help publishers, schools and companies distinguish AI-assisted material from content presented as entirely human-made. A machine-readable mark is potentially more useful than a detector guessing from vocabulary, sentence length and stylistic habits.

Others saw a loss of control. The strongest objections came from people who pay for Claude and do not want an invisible identifier placed in their private or professional work without an opt-out. The concern is not that a reader will notice the mark today. It is that a future employer, university, client or platform could use a detector to make a decision about them.

That concern is especially understandable for people who use AI as an editor rather than as a ghostwriter. A person can bring the experience, argument and original reporting, then ask Claude only to improve the language. If the result is later flagged as “AI-generated”, the tool’s modest contribution may overshadow the human work behind it.

Another line of criticism questions durability. Public commentary collected by Explainx focused on the lack of technical details, the missing detector and the possibility that paraphrasing could weaken the signal. The Register likewise reported scepticism among Claude users about whether text watermarking can deliver dependable proof. These are not merely arguments between people who like or dislike AI. They are questions about evidence.

Europe is setting the rule, but users will live with the result everywhere

Anthropic’s global rollout is significant for smaller European companies. They will not need to build one disclosure workflow for EU customers and another for clients elsewhere. It also means that a freelancer in Prague, a teacher in Finland and a developer in Canada may receive similarly marked output from the same supported model.

For an organisation using Claude in its own product, the responsibility does not end with the vendor’s mark. Anthropic explicitly tells developers to assess independently what Article 50 requires of their own services. A company that places Claude behind its own chatbot, document editor or customer-support system still needs to consider how it informs users and how it records AI involvement.

That is the practical European lesson: compliance is not the same thing as accountability. A mark can help explain that a system was involved, but it cannot replace a clear policy about who reviewed the result, who approved it and who is responsible for mistakes. Under GDPR, organisations should also be careful about turning provenance signals into unnecessary profiles of employees, students or customers.

What users should do now

There is no reason for ordinary Claude users to panic or to rewrite their entire workflow. The marks are intended to be invisible, and Anthropic has not announced a setting that users can inspect or switch off. The sensible response is to keep records of how AI was used when the context matters: drafting, translation, proofreading and original generation are not the same activity.

Schools and employers should wait for Anthropic’s technical documentation before adopting automatic penalties. They need to know how long a passage must be, how often the detector is wrong and whether independent parties can verify a result. Until those details exist, a mark should be treated as a prompt for a conversation, not as a verdict.

We have already covered the broader meaning of Article 50 transparency rules. Claude’s announcement shows what that principle looks like in everyday use: not a visible stamp on a screen, but a provenance signal hidden in the material itself. Whether that makes the digital world more honest will depend less on the existence of the mark than on the caution with which people interpret it.