Claude Now Watermarks AI Text. Here's Why You Still Need a Detector.

Quick answer

On August 11, 2026, Anthropic confirmed that every Claude model launched on or after August 2 now weaves an invisible statistical watermark into the text it generates, worldwide, across every Claude product. It's the first frontier lab to ship text watermarking at this scale. It is also, by Anthropic's own admission, not proof of anything on its own. The mark signals that Claude touched the text. It does not confirm Claude wrote it, and it disappears the moment someone paraphrases the output. Anthropic hasn't even shipped a way to check for the mark yet. Detection tools that read the statistical fingerprint AI writing leaves behind still do the job watermarking can't: work today, on any model, without waiting for a company to release a checker.

What Anthropic actually announced

Anthropic updated its Help Center on August 11 to explain two distinct marking systems. When generating text, supported Claude models embed a watermark directly into the text at the moment of creation. That watermark keeps moving with the text whenever you copy and paste it elsewhere. In some cases it still shows up after editing. For image files like SVGs, PNGs, and JPGs, Claude instead adds signed provenance data. This provenance data follows the C2PA standard, the same open framework that OpenAI committed to for images back in May.

With the EU AI Act's Article 50 now enforceable as of August 2, 2026, generative AI providers must mark their outputs in a machine-readable way. Anthropic went beyond the bare legal requirement: it applies the watermark to every region, not just EU traffic, across the Claude Platform API, claude.ai, Claude Code, Claude Cowork, and Claude Tag. Models that shipped before August 2 don't have the watermark yet. Anthropic says support for those is coming, with a target date tied to the EU's own transition deadline of December 2, 2026.

This closes a gap we flagged in our own watermark explainer back in late July: at the time, text watermarking existed only as an open research technique (Google's SynthID) with no major lab shipping it in a live text product. Claude is now the first to change that. The image-and-file side of the provenance story (SynthID, C2PA) hasn't moved much since then. The text side just did.

The mark says "processed," not "written"

Anthropic's own documentation is unusually direct about this limitation: "Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source."

Watermarks cannot act as detectors. The email is drafted by a person. Claude processes the draft, polishing its grammar. The polished version shows a Claude signature, suggesting Claude produced the entire content. A supervisor or instructor who spots the signature may conclude that the student or employee was not the author. That signature only indicates Claude's role in the workflow. It reveals nothing about how much of the content came from the human versus Claude. Consequently, authorship assessment should look at how much the human actually contributed.

Detection models built to read statistical patterns in finished text, perplexity, burstiness, the structural fingerprints different models leave behind, work from a different signal entirely. They don't need Claude, or any vendor, to have marked anything first. That's also their known limitation: a probability score, not a definitive verdict, the same honest-broker distinction we've made about our own detection scores. Neither approach alone proves authorship. Used together, a watermark check plus a detection score gives more signal than either does alone, which is exactly why the arrival of watermarking doesn't retire detection. It complements it.

Paraphrase breaks the mark. Detection still catches the result.

The bigger problem for anyone hoping a watermark settles disputes: it doesn't survive rewriting. The underlying technique traces back to a 2023 paper by John Kirchenbauer and colleagues at the University of Maryland, which biases a model's token choices toward a pseudorandom "green list" subtle enough that a reader never notices, but detectable with a statistical test over enough text. Google DeepMind's SynthID-Text, published in Nature in 2024, refined the same idea. Both share the same weakness: run the output through a paraphrasing pass, human or AI, and the biased token distribution the watermark depends on gets scrambled. The signal doesn't degrade gracefully. It's gone.

Anthropic does not dispute the claim. In an independent technology review, the announcement was described as having "what the mark cannot survive... is heavy paraphrasing." Anthropic's limitations page lists the same failure mode verbatim. Short texts suffer from a parallel problem. A statistical test needs about one hundred tokens or more to deliver a meaningful outcome. A reply consisting of only two sentences, or a simple social media post, fails to provide a trustworthy mark.

Even when watermarking loses its signal, content-level detection still succeeds. The trick lies in examining finished text for AI-specific patterns, independent of watermark survival. The detector's job is to read exactly what is shown on the page. Consequently, the failure of watermarking does not undermine its validity. Watermarking and content-level detection serve separate purposes. The challenge of effective content-level detection remains open. Anthropic has not published a public checker for its watermark.

No detection tool exists yet. That's the actual gap right now.

Here's the part that matters most for anyone who needs an answer this week: Anthropic has not released a way for anyone, developers, employers, schools, to check whether a piece of text carries its watermark. The company says only that it will "share details on detection mechanisms in forthcoming technical documentation." Until that ships, the mark exists, but nobody outside Anthropic can read it.

No practical value exists in the watermark for determining AI authorship today. The absence of a Claude watermark carries no significance. Text produced before August 2, from a model without watermark support, or that has been edited enough to lose the watermark signal will all return unmarked, no matter if they are AI-generated. Assuming human authorship because a watermark is missing would be the wrong approach.

None of this is a critique of Anthropic. A watermark that survives copy-paste is a real signal. Its value increases once the detection tool actually ships and enough labs adopt it that checking becomes routine. Still, the watermark doesn't replace what already works today: reading the text itself for the patterns AI writing tends to leave behind, no matter which model wrote it or whether that model's maker has gotten around to watermarking yet.

FAQ

If I ask Claude to edit my writing, does it now carry a watermark? Almost certainly, if you're using a Claude model released on or after August 2, 2026. Anthropic is explicit that the mark doesn't distinguish between Claude writing something from scratch and Claude lightly proofreading a human's draft. Both come back marked the same way.

Can someone remove the Claude watermark? Removal of the Claude watermark can be achieved. By applying a degree of paraphrase that exceeds the level needed to maintain the original statistical profile, the mark loses its detection advantage. Anthropic admits this plainly, refusing to pretend the watermark is impervious.

Does this mean text detectors are becoming unnecessary? No, and largely for the opposite reason people might expect. Detection works on the finished text itself, which means it doesn't depend on a specific vendor's watermark surviving a rewrite, or on that vendor shipping a public way to check for it in the first place, something Anthropic hasn't done yet for its own mark.

Is this watermark the same as the SynthID image watermarking OpenAI and Google use? No. SynthID and Anthropic's C2PA metadata both cover images and files. Claude's new text watermark is a distinct method built specifically for generated text. It uses the same class of token-biasing method that SynthID-Text and the original Kirchenbauer research established, just applied to a different model family.

Wondering what your own writing actually scores, watermark or no watermark? Try the free demo and see what the patterns say.