Can Claude's "Invisible Watermark" Be Removed?—A New Controversy in the AI Era Presented by Cardano Founder Anthropies

Distinguishing between AI-generated text and human-written text.

At first glance, this seems like a welcome mechanism in an era overflowing with fake information and a large volume of AI-generated content. However, if this "mark" is invisible to humans and could remain even in texts that have only slightly used AI, the issue is not so simple.

In August 2026, Charles Hoskinson, known as the co-founder of the cryptocurrency Cardano, released an open-source project called "Anthropies" to remove watermarks and provenance information related to Anthropic's generative AI "Claude."

This is more than just the release of a convenient tool; it raises larger questions about how much transparency should be technically enforced in generative AI and to what extent AI companies have the right to leave "marks" on content created by users.

Claude's "Invisible Watermark"

Understanding the mechanism of the text watermark introduced in Claude is crucial to grasping the current uproar.

When people hear "watermark," they might imagine logos displayed in the corners of images or secret strings embedded within text.

However, the method adopted by Anthropic is significantly different from that.

No special characters are added to the text. Instead, a statistical pattern is embedded in the "choice of words" itself when the AI generates text.

For example, when continuing a sentence, there may be multiple candidates like "important," "crucial," or "essential," any of which would not seem unnatural.

While a typical generative AI selects the next word based on a certain probability, the watermark method uses a rule based on a secret key to impart a statistical tendency to that choice.

Humans cannot discern the difference when reading the text.

However, those with the key can statistically evaluate the "possibility that Claude was involved in generating this text" by analyzing a sufficiently long text.

According to Anthropic, this method is based on the approach of Google's DeepMind's SynthID Text.

The important point is that this mark is not "data hidden behind the characters" but is embedded in the wording of the text itself.

Simply deleting special characters or removing metadata from a text file will not eliminate it.

The EU AI Act in the Background

One major reason Anthropic is introducing such technology is the European Union's AI regulation, the "EU AI Act."

The transparency provisions, effective from August 2, 2026, require providers of generative AI to implement measures to make AI-generated or manipulated content detectable by machines.

The scope includes not only text but also images, audio, and video.

With the rapid improvement in generative AI performance, distinguishing between human-created content and AI-generated content has become difficult. As a countermeasure against deepfakes, fake news, impersonation, and mass-produced promotional content, a system to verify the origin of content is being established.

However, the system is not as simple as "everything that has been slightly touched by AI must be displayed as AI-generated."

The European Commission's guidelines show exceptions for standard editing assistance, and source code, among other things, can be exempt from marking obligations under certain conditions.

Moreover, for systems that existed in the market before August 2, 2026, a limited transition period until December 2, 2026, is provided for marking and detection obligations.

Thus, it is not accurate to understand that "the EU has mandated watermarks for all AI usage."

Nonetheless, it is certain that the direction of making "AI-generated products mechanically identifiable" is strengthening among major AI operators.

"Anthropies" Released by Hoskinson

That's where Anthropies came into play.

Hoskinson announced on his X account that he is creating a new system that can be used in combination with many LLMs to remove Anthropic's watermark.

The project is published on GitHub and adopts the Apache 2.0 license.

The "traces of Claude" that Anthropies aims to handle can be broadly classified into three types.

The first is the statistical watermark embedded in the text itself.

The second is the Content Credentials in C2PA format attached to files like images.

The third is attribution information such as "Co-Authored-By: Claude" added to Git commit messages used in software development.

These three types may look similar but are technically entirely different in nature.

C2PA information and Git attribution are relatively easy to process as they exist as specific metadata or text.

On the other hand, the statistical watermark exists in the word choice of the text itself, so simply "removing the watermark part" is not possible.

Why "Rewrite with Another AI"?

The most intriguing aspect of Anthropies is its approach to text watermarks.

If the watermark exists as a pattern of word choice, then altering that pattern requires changing the text itself.

Therefore, Anthropies adopts the idea of reconstructing the original text using another model.

The project separates the process of removing known metadata and changing the structure and expression of the text.

For Git attribution, it is sufficient to remove the designated parts, but for natural language text, it is necessary to significantly rewrite it with a model different from the one that generated the watermark.

Conversely, if you ask the same model that generated the watermark to "erase this watermark and rewrite it," there is a possibility that the newly generated text will re-embed the same type of watermark.

This reflects the fundamental difficulty of watermark technology in generative AI.

Is There Meaning in a "Removable Watermark"?

This is precisely the question frequently seen on social media.

"If the text can be largely paraphrased by another AI to remove it, can it really prevent malicious use by humans?"

Indeed, Anthropic itself does not describe the watermark as foolproof.

While mild editing might leave statistical features, a complete rewrite could potentially erase the signal.

Additionally, short texts have fewer word choice instances that serve as judgment materials, leading to decreased detection capability compared to long texts.

The same applies to program code or texts describing strict facts.

In places where there is almost only one correct answer, AI has little room to freely choose different words, reducing the places where statistical patterns can be embedded.

On social media, there are criticisms like "If the provenance information of images can be lost through screenshots or re-encoding, and the signal of text can be weakened by extensive paraphrasing, it may not be effective against truly deceptive users, only affecting ordinary users."

On the other hand, there is also the opinion that "Even if it cannot prevent all misuse, it still has meaning."

Not all AI-generated content posted en masse on the internet is necessarily processed with watermark removal in mind.

Even being able to detect cases where the generated result is almost copied and posted can serve as a means to distinguish a large volume of spam or low-quality content.

The Noticeable Concern on Social Media: "What About Human-Written Texts?"

Another strong reaction on social media is about the treatment of "texts that were written by oneself and proofread by Claude."

For example, suppose a report written entirely by a human is given to Claude to refine the expressions for readability.

If a statistical pattern emerges in the parts changed by AI, there is a possibility that it will later be judged as "a text involving Claude."

The important point here is that the judgment of "involvement of Claude" is not the same as the judgment of "Claude wrote this text."

Anthropic also makes this distinction clear.

What the watermark reveals is the possibility that Claude was involved in the text, not definitive evidence of whether the author was human or AI.

Still, if the judgment result alone is taken out of context in schools, companies, recruitment, or publishing, it could pose a problem for users.

On social media, there is a growing reaction that it is unacceptable for a text written by a human to be considered "AI text" just because it was run through Claude for copy editing purposes.

However, Anthropic explains that in texts where only grammar or punctuation is lightly corrected, there are often not enough word choices by Claude for a sufficient watermark to remain.

The EU guidelines also provide exceptions for standard editing assistance in marking obligations.

Thus, it is not the case that "if even one character is corrected by Claude, it will always be judged as AI-generated."

The problem is that the boundary between "standard editing" and "substantial generation by AI" can easily become ambiguous in real-world usage scenarios.

Voices Welcoming Transparency

Of course, there are also many opinions that positively evaluate the watermark itself.

On social media, there are opinions that a mechanism to technically verify AI-generated content will be necessary for the future of the internet.

As AI-generated texts, images, and videos approach a level indistinguishable from human creations, the means for viewers to judge the source of content becomes important.

Considering political misinformation, fake advertisements posing as companies, fictitious news, fake reviews, and a large volume of SEO content automatically generated by AI, the technology to confirm the "possibility of AI generation" has public significance.

If AI companies do nothing voluntarily, the traceability of generated products will further decline.

In that sense, some users perceive Anthropic's efforts as a "step forward in enhancing transparency."

Some posts suggest that even if the detection technology is not perfect, being able to distinguish AI content that is published en masse without processing could lead to an improvement in the overall information quality of the internet.

The "Ownership" Issue Highlighted by Hoskinson

Viewing Anthropies merely as a technical watermark removal tool only partially understands the project's intent.

Hoskinson's side strongly emphasizes the issues surrounding the "ownership" and "attribution" of AI-generated products.

Anthropies' documentation questions the continuous presence of AI operator traces in products owned by users.

In particular, displays like Git's "Co-Authored-By: Claude" differ from mere mechanical provenance information and may appear to humans as if Claude is displayed as a co-author.

However, it is important to note that this Git attribution and the new statistical watermark are separate mechanisms.

The Anthropies repository itself explains that Git's co-author attribution existed before the current watermark system and that there is no evidence Anthropic positions both as the same mechanism.

Anthropic has also made it clear that the watermark does not indicate "ownership or authorship."

Thus, at this point,

"The presence of a watermark could potentially lead to ownership risks in the future"

is a concern raised by Hoskinson's side,

while "The watermark does not change user rights"

is Anthropic's stance.

These two need to be considered separately.

The "AI Provenance War" Has Only Just Begun

The emergence of Anthropies may have highlighted a larger structural issue rather than the technology itself.

AI companies strengthen watermarks.

Then, technology to remove them emerges from the user side.

AI companies improve detection technology.

Another tool then reconstructs the text to avoid detection.

If this trend continues, it could become a "cat-and-mouse game" over content generation and detection.

Indeed, after the announcement of Claude's watermark, posts on X quickly began exploring removal methods and the effects of paraphrasing, and services and open-source tools claiming to remove watermarks have started gaining attention.

Anthropies has become a symbol of this trend.

The Question Is Not "Did You Use AI?" but "How Did You Use It?"

In the future, what will become more important is not simply classifying content into "human-made" and "AI-made."

Cases where humans plan, AI drafts, and humans completely rewrite.

Cases where humans write everything, and AI only corrects typos.

Cases where AI-generated text is fact-checked by humans, who then add interview information and edit it.

Or cases where AI writes articles 100% automatically and publishes them en masse.

Current generative AI usage forms such a continuous spectrum.

The "involvement of AI" detected by the watermark does not necessarily match the "who is responsible for the content" that society truly wants to know.

On the other hand, without any means of identification, it becomes difficult to distinguish fake information or spam mass-generated by AI.##HTML_TAG_257