Anthropic explains how Claude’s new watermarks will work — and what users can still do to remove them

Anthropic has begun explaining how it will watermark text generated by Claude, a move designed to satisfy European Union transparency rules and reassure users that the system will not change the quality of its output. The company says the markings will be invisible to readers, detectable by approved tools, and difficult to erase without substantially rewriting the text.

The clarification matters because the policy has already triggered a backlash among some Claude users, including accusations that the feature could expose people using AI for work or school. Anthropic is now trying to show that the system is meant to identify machine-generated content, not degrade the writing itself.

Why Anthropic is adding watermarks to Claude

Anthropic is introducing text watermarking to comply with the European Union AI Act’s transparency requirements. The company has said the new rules require AI providers to make AI-generated material identifiable, and watermarking is one way to do that without making the output look different to the average reader.

The move places Anthropic alongside other large model developers that have signed the same Code of Practice and are expected to deploy their own identification systems. In practical terms, this is one of the clearest signs yet that watermarking is moving from a research idea into a standard product feature for major AI chatbots.

What the rule is trying to solve

The EU’s transparency framework is aimed at making synthetic content easier to identify in a world where text, images, audio and video can be generated at scale. For chatbot makers, the challenge is to create a signal that survives normal use, but does not affect how the output reads or behaves.

Anthropic’s response suggests the company believes text watermarking can meet that bar, at least for ordinary conversational use. The company also appears to be drawing a line between identifying AI-authored text and policing whether people use AI assistance at all.

How does Claude’s watermarking actually work?

Anthropic says the system works by making subtle word choices when Claude is generating text. In situations where there are several acceptable options — for example, describing weather as “overcast” rather than “grey” — the model can follow a pattern that embeds a hidden signal in the response.

That signal is not meant to change meaning or style in a way readers would notice. Instead, it is designed to be readable only by someone or something with the right key. In other words, the text should look normal, but carry a machine-detectable mark underneath the surface.

Anthropic says the watermark is meant to be invisible to readers while remaining detectable to authorized tools that know how to verify it.

The company also emphasized that watermarking should not alter the quality of Claude’s output. According to Anthropic, readers will not be able to tell the difference between a watermarked response and one that has no watermark at all.

What technology is Anthropic using?

Anthropic says it will use the SynthID-Text approach, the watermarking method originally outlined by Google DeepMind in 2024. The company also plans to release a watermark detection API so developers and other users can check whether a piece of text was generated by Claude.

That choice is important because it connects Anthropic’s rollout to a broader industry effort rather than a one-off proprietary experiment. By using a known watermarking framework and offering a detection tool, Anthropic is signaling that the feature is meant to be interoperable and practical in real-world settings.

Can people edit the watermark away?

Yes, but Anthropic says simple edits probably will not eliminate it. The company’s explanation is that the system is robust enough to survive light revisions, but not a wholesale rewrite that replaces every word.

That distinction is central to the usefulness of the tool. If a watermark disappears whenever someone changes a few words or corrects grammar, it would be of little value. If it remains after ordinary editing, it becomes much more useful for verifying whether the original text came from Claude.

What counts as light editing?

Anthropic did not define the term with a formal threshold, but it suggested that minor proofreading or small wording changes would not usually remove the mark. In that case, the company says, the underlying text still contains most of the original model-generated structure.

By contrast, a full rewrite would erase the watermark because the original word pattern would be gone. Anthropic went further and noted that if every word has been replaced, the resulting passage may no longer be fairly described as AI-generated content in the first place.

Anthropic says a complete rewrite can remove the watermark, but argues that at that point the text has effectively become a human-authored version rather than a direct AI output.

What about text Claude helped edit, not write from scratch?

That case is more complicated. Anthropic says detectability will depend on the length of the material and the extent of Claude’s revisions. If the model only lightly edits a human draft, the company says nearly all of the final words may still come from the human author, leaving little for the watermark to attach to.

In other words, the more Claude is used as a proofreading or rewriting assistant rather than a full drafting engine, the less reliable the watermark may become as a signal of AI generation.

What happens to code?

Code should carry less of a watermark than ordinary prose, Anthropic says, because programming is constrained by functionality. A model writing code cannot freely swap between many equivalent terms the way it can in natural language, since the output must remain syntactically and logically correct.

That means the system has fewer opportunities to hide a pattern in the wording. Still, Anthropic says some parts of code do include arbitrary wording, especially comments, and those sections can be watermarked. The company says the overall effect on actual code should be negligible.

This is an important limit for developers who use Claude as a coding assistant. It suggests Anthropic is trying to avoid interfering with the correctness or readability of software while still preserving some level of traceability in ancillary text.

How is this different from AI detection tools?

Anthropic is drawing a sharp distinction between watermark detection and AI-style detection. Watermarks are embedded at generation time and can be checked with the right key or API. Detection tools, by contrast, infer whether text looks AI-written by searching for patterns, repeated constructions or other stylistic clues.

That difference matters because pattern-based detectors can be wrong. A human writer may sound “AI-like,” while a model-generated paragraph can be edited enough to look fully human. Anthropic says watermarking is a more direct and reliable method because it verifies a signal inserted by the model itself.

The company says identifying stylistic fingerprints in a passage is fundamentally different from checking for a deliberate watermark embedded during generation.

Why this distinction matters to users

For employers, educators and content reviewers, the difference between “looks AI-generated” and “contains a watermark” is significant. A watermark can serve as stronger evidence that a piece of text came from a model, while a detector that guesses based on style may be less dependable and more controversial.

For users, the distinction also affects fairness. Some Claude users fear that watermarking could be used to penalize them for relying on AI tools in situations where they are allowed to do so. Anthropic’s framing suggests the company sees the feature as a transparency mechanism, not a disciplinary one.

Why are some Claude users upset?

Because many users interpret watermarking as a threat to privacy, productivity and flexibility. After Anthropic first disclosed the plan earlier in the week, some users on Reddit argued that the feature was aimed at innocent users rather than bad actors. Others pushed back, claiming that only people trying to mislead others would object.

Reactions on X have been equally heated, with Business Insider reporting that dozens of users said they had cancelled or planned to cancel subscriptions. Whether that backlash proves lasting or temporary remains unclear, but it shows how sensitive AI transparency measures have become among people who rely on chatbots for everyday work.

The emotional intensity around this feature also reflects a broader unease in the AI market. Users increasingly want powerful tools, but they also do not want those tools to create permanent records of how they work, study or communicate.

What this means for Anthropic and the AI industry

Anthropic’s watermarks are about more than one chatbot setting. They point to a future in which major AI systems are expected to build provenance and traceability into the products themselves. That future could reshape publishing, education, enterprise workflows and legal compliance.

For Anthropic, the rollout may help reduce regulatory risk in Europe while signaling to businesses that Claude is being positioned as a responsible, enterprise-ready platform. At the same time, it may create friction with some customers who use AI in ways they would prefer not to disclose.

There is also a competitive angle. If other major model makers follow the same route, watermarking could become a baseline feature rather than a differentiator. But if some vendors resist or implement weaker versions, users may gravitate toward the systems that are easiest to hide or hardest to trace.

Potential winners and losers

• Regulators gain a clearer path to enforcing transparency requirements.
• Employers and schools may get a better way to verify AI-assisted text.
• Power users could face more scrutiny around how they use Claude.
• Anthropic may improve compliance but risk alienating some customers.
• Other model developers may be pushed to adopt similar systems.

Timeline: How Claude’s watermarking story unfolded

The debate over Claude’s watermarking did not start with Friday’s clarification. It has been building since Anthropic revealed earlier in the week that the feature was coming.

How should users think about the new system?

Users should think of the watermark as a transparency layer rather than a visible label. It is meant to preserve the usefulness of Claude while giving institutions a way to verify origin when needed.

That said, the feature is unlikely to end the dispute over AI disclosure. People who want to hide AI assistance may see the watermark as intrusive. Institutions that want accountability may see it as overdue. Anthropic is trying to satisfy both sides, but the company’s latest explanation makes clear that the trade-offs are real.

What happens next may depend less on the technical details than on whether users, businesses and regulators come to trust the system. If the watermark can reliably identify machine-generated text without creating obvious drawbacks, it may become a model for the industry. If not, it could become another reminder that transparency tools are easy to announce and hard to normalize.

For now, Anthropic is betting that invisible marks, detection APIs and selective limits on code are enough to make Claude compliant without making it less useful. The company’s challenge is proving that promise in a market where every new AI safeguard tends to be read as either a breakthrough or a threat.