A Compliance Move Becomes a User Debate

Anthropic has published a new blog post explaining how it plans to watermark text generated by Claude, addressing three practical questions users have been asking: how the watermark works, whether editing can hide it, and what happens when Claude writes code.

The move follows Anthropic’s earlier disclosure that it would introduce watermarking to comply with the EU AI Act’s Transparency Code, which requires AI companies to use systems that make AI-generated content identifiable. The announcement has divided Claude users. Some Reddit users portrayed the change as hostile to ordinary customers, while others argued that resistance to watermarking is mainly about hiding AI use. Business Insider also reported that “dozens” of users on X claimed they were canceling Claude subscriptions over the plan.

How the Text Watermark Works

Anthropic describes the watermark as an invisible statistical pattern rather than a visible label. When Claude has low-stakes choices—such as choosing between words like “overcast” and “grey” to describe weather—it can make choices that encode a pattern. Readers should not notice the difference, but someone with the appropriate key can detect the watermark.

Anthropic says the watermark does not reduce Claude’s output quality and that a watermarked response should look the same to a reader as an unwatermarked one. The company said it will use the SynthID-Text approach described by Google DeepMind in 2024 and plans to release a watermark detection API.

The company also distinguishes watermarking from conventional AI text detection. Many AI detectors look for stylistic “tells” in writing, such as repeated sentence structures or common rhetorical patterns. A watermark, by contrast, checks for a deliberately encoded signal in the generation process.

Key points include:

• Regulatory driver: compliance with the EU AI Act’s Transparency Code;
• Technical approach: SynthID-Text;
• Planned tooling: a watermark detection API;
• Broader industry shift: Anthropic says other major model developers that signed the same Code of Practice will implement their own watermarks.

Editing, Rewriting, and Claude-Assisted Proofreading

Anthropic says light editing probably will not remove the watermark completely. A full rewrite that replaces every word can remove it, but the company notes that, at that point, it becomes debatable whether the final text should still be described as AI-generated.

That means the watermark is not an unbreakable lock. It is better understood as a signal designed to survive ordinary use of Claude’s output, not a guarantee against deliberate rewriting.

For text that Claude only proofreads or lightly edits, the answer depends on length and the degree of editing. If a human wrote nearly all the words and Claude made only minor changes, Anthropic says there may be little or nothing for the watermark to attach to.

Why Code Is a Special Case

Code will carry less watermarking than ordinary prose, according to Anthropic, because the model has less freedom when generating working software. Programming output must satisfy syntax and functional requirements, so Claude cannot freely swap equivalent words the way it can in natural language.

Still, Anthropic says watermarking may appear in places where arbitrary choices exist, such as comments inside code. The company says any effect on actual code should be negligible.

For developers, the practical takeaway is that watermarking is not presented as a change to Claude’s coding capability. The bigger concerns remain code correctness, review, security, and compliance with relevant licenses.

The Bigger Direction for AI Transparency

Anthropic’s explanation shows how AI governance is moving from broad principles into product-level mechanisms. Watermarking will not solve every provenance problem: rewritten text can lose the signal, and lightly edited human writing may not contain enough Claude-generated material to detect. But it gives regulators, platforms, and organizations a concrete way to begin identifying AI-generated content.

The next test will be implementation. Detection reliability, clear limits, and responsible use of results will matter as much as the watermark itself. If treated as one signal among several, watermarking could become part of standard AI transparency infrastructure; if treated as definitive proof in every case, it is likely to remain controversial.