AI Transparency & AI Governance

Claude Introduces Invisible Text Watermarks: Why the AI Act Deserves Our Thanks

Anthropic is introducing machine-readable text watermarks for Claude. This statistical technique works without invisible characters or additional metadata. It’s also important to consider why we’re even hearing about this.

The Geviertstrich can breathe a sigh of relief.

A whole set of superstitions has now developed around the detection of AI-generated texts. Four dashes, certain sentence rhythms, unusually correct phrasing, or suspiciously neatly structured lists of three are quickly deemed proof that a text must have been generated by artificial intelligence.

That's not very durable.

Anthropic is therefore taking a technically much more interesting approach with Claude. Future Claude models will embed a subtle, machine-readable watermark in generated text from the outset; existing models will be retrofitted gradually. Anthropic actually describes the process as Claude “weaving” the mark directly into the text .

So are there secret, spiderweb-like structures between our words after all?

Only in a figurative sense.

Update from August 15, 2026:

On August 14, Anthropic provided further details on the technical implementation of the Claude watermark. The method is based on a variant of Google DeepMind’s SynthID-Text. Neither invisible characters nor additional metadata are inserted into the text. Instead, a key-driven random process influences the selection among several words that are appropriate in terms of content. Over longer passages of text, this creates a statistically recognizable pattern.

The watermark does not contain any information about users, organizations, or individual chats, and does not require any additional tokens. However, for short texts that are highly factual or have been edited only minimally by Claude, the signal may be too weak for reliable detection. A public detection API has been announced but is not yet available.

Particularly noteworthy is the now explicit rationale: Anthropic explains that it is introducing the watermarks to comply with the transparency requirements of the EU AI Act. The central argument of this article is thus confirmed by the provider itself.

A close-up of a spiderweb covered with water droplets, symbolizing traces of origin woven invisibly into AI-generated texts — A spiderweb with fine water droplets symbolizes machine-readable structures that Claude weaves into generated texts.

Invisible traces of origin are not a new invention

The principle of embedding information about a document's origin—information that is barely perceptible to humans—is much older than generative AI.

It has been documented since the 2000s that certain color laser printers and color copiers leave tiny yellow dots on every page. The Electronic Frontier Foundation was able to decipher a repeating pattern of 15 by 8 dots on devices from the Xerox DocuColor series.

The dot pattern included, among other things, the printer's serial number and the date and time of the printout. Under normal lighting, these dots were barely visible. However, when magnified or viewed under blue light, they became discernible and could then be analyzed.

A printed document thus contained an additional layer of information that most users were unaware of.

The historical parallel with Claude is intriguing: The AI-generated text, too, bears a trace that remains imperceptible to the reader but is intended to be recognized by machines.

Technically, however, the two methods differ fundamentally. A color laser printer actually adds additional dots to the document. Claude, on the other hand, adds neither invisible Unicode characters nor any other characters or metadata to the text. The watermark is created during the generation process through a statistically influenced selection of appropriate words.

The level of detail is also different. With the decrypted printer patterns, it was possible to identify a specific device and the time of printing. According to Anthropic, however, the Claude watermark contains no information about users, companies, accounts, or individual chats. Its sole purpose is to indicate that a piece of content may have been processed by Claude.

How Claude Labels AI-Generated Content

Anthropic relies on two different techniques.

1. Embedded Watermarks for Text

When a supported Claude model generates text, a variant of Google DeepMind’s SynthID text is used. A secret key influences the random selection among several contextually appropriate next words or tokens. Over the course of a longer passage, this creates a statistical pattern that can be verified using the appropriate detection method and key.

"Invisible" here explicitly does not mean that Claude inserts hidden Unicode characters, appends data to the text, embeds text metadata, or always uses the same suspicious phrasing. The spiderweb metaphor therefore describes only the distributed structure of the signal, not its technical implementation.

Since the markup is generated from the text itself, it should be preserved when copied and pasted and should be able to withstand minor edits in some cases. The markup is applied at the model level and is therefore not limited to the standard Claude interface.

Supported models are expected to use the watermarks in Claude, via the API, in Claude Code, Claude Cowork, and on various cloud platforms, among other places. New Claude models introduced in the EU since August 2, 2026, support watermarking from the start. Anthropic is still working on a retrofit for older models.

This also means that, at this time, one should not assume that every text—whether current or ever generated by Claude—already contains a recognizable watermark.

According to Anthropic, the process does not require any additional tokens and therefore does not incur any additional token-based usage costs. The impact on generation speed is said to be negligible.

2. Signed provenance metadata for files

For supported file formats such as SVG, PNG, and JPG, Claude adds digitally signed provenance metadata.

These follow the open C2PA standard. C2PA is designed to document information about the origin and processing of digital content. A valid signature can indicate that a file was processed by Claude. It is also intended to make it possible to determine whether the file has been modified subsequently.

This is a different approach than the text watermark. With text, the signal is embedded in the output itself. With files, a signed information layer is also included.

However, metadata is not indestructible. Format conversions, resaving, or screenshots can remove such information. This is another reason why Anthropic does not rely exclusively on traditional file metadata for text.

We now know the procedure, but not yet the specific method of detection

With SynthID-Text, we now understand the basic process: The probability distribution of possible next tokens is modulated based on a key in such a way that a statistical signal emerges over longer passages.

However, important details regarding the specific implementation of Claude remain unclear. Anthropic has announced a detection API but has not yet made it publicly available. Details regarding key management, detection thresholds, and the practical error rate are also still missing.

We are therefore not yet able to independently verify the Claude text watermark or reproduce the detection.

That's an important point.

A provenance signal only realizes its full value when independent systems can reliably detect it and correctly interpret the results. Otherwise, a technically interesting feature remains, for the time being, merely a claim made by the provider.

The presence of a watermark does not prove that the work was created by AI

The greatest danger is inferring more from the label than it actually implies.

According to Anthropic, if a Claude tag is detected, it simply means that the content may have been processed by Claude.

The original content may still have been created by a human. For example, a detectable signal can arise when Claude generates the text:

translated extensively,
In summary,
substantially rewritten,
restructured
or have converted it to another file format.

According to Anthropic, however, a simple spelling or grammar check often yields hardly any reliable, discernible signal, because Claude selects only a few words on its own. A translation generated entirely by Claude can be flagged much more strongly, since the model regenerates nearly every word.

Therefore, the discovery of a note does not prove that Claude developed the original idea or wrote the entire text on his own.

Similarly, the absence of a mark does not prove human authorship.

The flag may be missing if the text was generated using an older model, completely rewritten, combined with other content, or significantly shortened. Minor edits may preserve the flag; a comprehensive rewrite is likely to remove it.

The content itself also affects how recognizable it is:

Short texts do not provide enough statistical data.
Passages that are heavily factual allow for fewer equally suitable word choices.
Program code usually carries a weaker signal than normal running text.
Proofreading alone often involves too little text to provide reliable evidence.

At best, the watermark answers the question:

Was Claude likely involved in processing this content?

It does not answer:

Who came up with the idea, who verified the statements, and who is taking responsibility?

This means that the watermark is an indicator of origin. It is not a lie detector, not proof of plagiarism, and certainly not a seal of quality.

We really should say thank you to the AI Act for once

The EU AI Act is often associated primarily with bureaucracy, documentation requirements, and additional obligations.

This criticism is not entirely unfounded. However, the current case shows that regulation can also provide very tangible benefits.

Article 50 of the AI Act has been in effect since August 2, 2026. Providers of generative AI systems must, as a general rule, design their systems so that artificially generated or manipulated audio, image, video, and text content is marked in a machine-readable format and can be identified as such.

This requirement does not apply unconditionally to every minor text edit. The EU guidelines list, among other things, standard edits as an exception when the system operates solely in a supporting role and does not significantly alter the input data or its semantics. This is consistent with Anthropic’s technical explanation of why simple proofreading often does not generate a reliable watermark signal.

The technical solutions implemented should, to the extent technically feasible, be effective, reliable, robust, and interoperable.

In addition, the EU has established a code of conduct on transparency for AI-generated content. Participation in this code is voluntary. The legal transparency requirements under Article 50 are not.

Anthropic has signed the code of conduct and has stated that it will implement machine-readable labels to promote transparency and comply with legal obligations.

That is why we must allow ourselves to ask an uncomfortable question:

If it weren't for the transparency requirement in the AI Act, would we even know that Anthropic weaves such structures into its texts?

There is no definitive answer to that question. One should not assume that Anthropic would have remained silent in the absence of regulation.

One thing is certain, however: Without Article 50, there would be no comparable legal pressure to systematically introduce such origin indicators, document how they work, and provide third parties with the means to identify them.

Regulation does not, in this context, ensure that AI-generated text will automatically be accurate or of high quality. However, it does ensure that providers must be at least somewhat more transparent about their technical methods.

For that, we really do have reason to thank the AI Act.

What Companies Should Take Away From This

Stylistic features are not an examination method

Quotation marks, certain phrasing, or perfectly structured lists are not reliable indicators of AI use.

Such characteristics can provide clues, but they inevitably lead to numerous misjudgments. Professional governance must not be based on whether a text sounds like it was generated by AI.

Origin and responsibility must be considered separately

A technical indicator may suggest that an AI system was involved. However, it cannot determine whether the content has been fact-checked, is legally permissible, or makes strategic sense.

This responsibility rests with the company and the individuals who approve the content.

Human testing must be substantial

For certain AI-generated or manipulated texts on matters of public interest, the AI Act requires operators to label them. An exception may apply if there is genuine human review or editorial oversight, along with associated responsibility.

According to the European Commission’s guidance, a simple spell-check or superficial formal review is not considered sufficient human verification.

Companies should therefore not only document which tool was used; they should also specify who reviews the content for accuracy and who bears final responsibility for its publication.

Source signals belong in content governance

For images, videos, documents, and—in the future—text as well, provenance information is increasingly becoming part of professional content workflows.

These include:

the collection of relevant metadata,
the verification of digital signatures,
clear approval processes,
Transparent editorial responsibilities
as well as guidelines for the use of generative AI.

Anyone who uses AI in marketing, communications, documentation, or customer processes should not wait until a regulatory agency or a customer asks about it before addressing these questions.

Transparency is infrastructure

The discussion about AI-generated content should move away from supposedly typical punctuation and phrasing.

What is more important are reliable traceability signals, open standards, transparent processes, and clear accountability.

The Claude watermark is not yet a perfect method. Its basic principle is now understood, but its specific detection and implementation cannot yet be independently verified. Markers can disappear, and their interpretation remains limited.

Nevertheless, this approach is far more credible than the claim that AI-generated texts can be reliably identified by the use of em dashes or particularly correct language.

The AI Act does not automatically make AI-generated content true, good, or responsible.

However, it forces providers to make something visible that might otherwise have remained invisible without regulation.

Claude can weave the technical details into a text.

Responsibility for its content remains with the individual.

What are you looking for?