Claude can now leave something behind when it writes for you—and no, it is not an em dash.

Anthropic has announced that supported Claude models will embed an imperceptible watermark directly into generated text. You will not see a label or strange symbol, but the watermark may stay with the text when it is copied, pasted, or lightly edited.

That has naturally led to some dramatic reactions online: Will teachers finally be able to catch every AI-written essay? Can publishers identify Claude-written books? Are Claude watermark removers about to become a thing?

The reality is a little less dramatic and much more complicated.

Does Claude watermark every response now?

Not quite.

Claude models released on or after August 2, 2026 will support watermarking from launch. Anthropic is also working to add it to older models, but it has not yet published a complete list showing which existing models are already covered.

The watermark will apply worldwide, not just inside the European Union. It will also follow supported models across Claude’s main app, API, Claude Code, cloud platforms, and other Anthropic products. Supported image files, including PNG, JPG, and SVG files, will receive signed provenance metadata based on the C2PA standard.

So the claim that every Claude response ever created is now watermarked would be inaccurate. Older outputs will not suddenly acquire a watermark, and some existing models are still being updated.

How does the Claude text watermark work?

Anthropic has not shared the exact technical method yet.

What it does say is that Claude “weaves” the watermark into the text at the model level. The mark does not visibly change the writing, its meaning, or its readability. It should also travel with the words when someone copies the text into a document, email, website, or publishing platform.

This is important because many websites claiming to remove “Claude watermarks” currently focus on deleting:

• Zero-width characters
• Strange Unicode symbols
• Formatting artefacts
• Em dashes or unusual punctuation

Those tools may clean up formatting, but there is no public evidence that they can remove Anthropic’s newly announced model-level watermark.

Until Anthropic releases its technical documentation, anyone confidently claiming to know exactly how the watermark is embedded is mostly guessing.

How can you check whether text has a Claude watermark?

At the moment, you cannot reliably check it yourself.

Anthropic says it is building detection tools for users and third parties, but those tools and the supporting documentation have not been released yet. The future checker will look for an official Claude mark rather than simply guessing whether the writing “sounds like AI.”

That makes it different from current AI-content detectors.

Traditional AI detectors examine writing patterns such as predictable wording, sentence structure, repetition, and token probability. They can flag human writing and miss AI writing because they are making a statistical guess.

A watermark detector would instead look for a signal intentionally placed there by Claude.

There is still one big catch: finding that signal would not necessarily prove that Claude wrote the content.

Does a Claude watermark prove that the text was AI-written?

No and this may be the most important detail in the whole announcement.

Suppose you write a complete article yourself and then ask Claude to fix the grammar. The edited version may carry Claude’s watermark even though the research, ideas, arguments, and original draft were yours.

The same could happen when Claude is used to:

• Translate human-written content
• Summarise an existing document
• Improve formatting
• Convert a file
• Rewrite one awkward paragraph
• Proofread a manuscript

Anthropic explicitly says a detected mark only indicates that the content may have been processed by Claude. It does not establish that Claude was the original author.

That distinction matters for schools, employers, publishers, and anyone considering using the detector as evidence of misconduct.

Think of the watermark as a receipt showing that Claude handled the text—not a confession proving that Claude created it.

Can the Claude watermark be removed?

There is currently no verified, one-click Claude watermark remover.

Anthropic does acknowledge that substantial changes can make the watermark harder or impossible to detect. Heavy editing, paraphrasing, translation, or combining Claude’s output with other writing may weaken the signal. Very short passages may also contain too little information for reliable detection.

But this creates an important two-way limitation:

A detected watermark does not prove Claude authored everything.

An undetected watermark does not prove Claude was never used.

So even after Anthropic releases its checker, it will not become a perfect test for separating human and AI writing.

Why is Anthropic introducing text watermarks now?

The immediate reason is the EU AI Act.

Article 50 of the Act became applicable on August 2, 2026. It requires providers of generative AI systems to add machine-readable marks that can help identify AI-generated or manipulated content. Anthropic has signed the EU’s related Code of Practice and is using watermarking to meet those commitments.

The regulation is European, but Anthropic has decided to apply the marks globally.

That does not mean every blog post, email, or caption edited with Claude must carry a large visible “AI-generated” label. The EU’s visible disclosure requirements are narrower and mainly focus on deepfakes and public-interest text published without meaningful human review or editorial control. Human-reviewed publications can qualify for an exemption from that labelling requirement.

The invisible mark and the visible disclosure label are two separate things.

What does Project Panama have to do with the Claude watermark?

There is no evidence that Project Panama caused Anthropic to introduce text watermarks. The legal deadline under the EU AI Act is the clear immediate reason.

Still, Project Panama gives the announcement an interesting and slightly uncomfortable context.

Project Panama was Anthropic’s effort to buy, cut apart, scan, and dispose of millions of physical books so their contents could be used in developing AI models such as Claude. The project became public through documents unsealed in a copyright case involving authors and Anthropic.

The connection is not technical. It is about two different sides of AI provenance:

• The Claude watermark helps identify what came out of the model.
• Project Panama raised questions about what went into the model.

A Claude watermark may tell us that the model handled a particular paragraph. It will not tell us which books, articles, or other human-created works influenced that paragraph, whether their creators gave permission, or whether they were compensated.

So Project Panama belongs in the broader conversation—but it should not be presented as the reason for the watermark.

What happens next?

The real test will begin when Anthropic releases its detection system.

We still do not know how accurate the checker will be, how much text it will need, which languages it will support best, how it will handle code, or whether public access will be unrestricted.

Until then, be cautious about tools claiming they can already detect or remove the official Claude watermark.

Anthropic has taken a meaningful step toward making AI involvement more traceable. But traceability is not the same as authorship, and a watermark should never be treated as a final verdict on who actually created the work.