Anthropic Is Watermarking Claude’s Text to Comply With EU Law. Here’s How It Works.

Every response Claude writes is about to carry a signature you can’t see.

Anthropic confirmed in mid-August 2026 that new Claude models will embed an invisible watermark into the text they generate — a statistical pattern woven into the model’s word choices, detectable only by someone holding a secret key. The company says the change has no effect on the quality, cost, or speed of Claude’s output. It also says there’s no way to turn it off.

The trigger is the European Union’s AI Act. Article 50 of the law, which requires AI systems to make synthetic content machine-detectable, took legal effect on August 2, 2026. Anthropic is one of roughly 190 organizations — including Google, Meta, Microsoft, Mistral, and OpenAI — that signed the EU’s Code of Practice on Transparency of AI-Generated Content, the voluntary framework regulators designed to help companies meet that requirement. Anthropic laid out the technical details in a blog post published August 14, following an initial support-page update a few days earlier that first alerted users to the change.

The rollout has produced two very different conversations. One is about whether embedding a watermark can degrade the writing itself — a concern raised almost immediately by outside critics. The other, which has generated far more visible user pushback, is about consent: the marking applies by default, worldwide, to everyone, with no way to opt out.

Why Now — the EU AI Act

The specific legal hook is Article 50 of the EU AI Act, which requires providers of AI systems that generate synthetic content to mark that output in a way machines can detect. The obligation became enforceable on August 2, 2026.

To help companies comply without each one inventing its own approach from scratch, the European Commission published a Code of Practice on Transparency of AI-Generated Content on June 10, 2026, and issued an opinion on July 8 finding that the Code adequately covers the relevant parts of Article 50. Signing it is voluntary. The underlying legal requirement is not — it applies to any provider whose AI output reaches the EU market, regardless of where the company is based.

Anthropic joined a large group of companies signing on, including Google, Meta, Microsoft, Mistral, and OpenAI. Notably, xAI has not signed the Code, though it would still be subject to Article 50’s binding requirements if it serves EU users.

How the Watermark Actually Works

To understand what’s actually happening inside Claude’s text generation, it helps to know how language models pick words in the first place.

A model like Claude generates text one word at a time. At each step, it’s choosing from a list of plausible next words. Often, several of those choices are roughly equivalent — in Anthropic’s example, a sentence like “the weather today was cold and ___” could reasonably end in “overcast” or “grey,” and it wouldn’t much matter to a reader which one the model picked. Normally, that kind of tie is broken by an arbitrary random number.

Watermarking changes where that randomness comes from. Instead of an arbitrary random number generator, the choice is guided by a secret cryptographic key combined with the preceding words. The model still ends up picking from the same pool of reasonable options — it isn’t nudged toward strange or unlikely words — but the sequence of choices it makes now follows a pattern that’s invisible to a reader, yet checkable by anyone who has the key. Enough of these choices accumulate across a passage to produce a statistical signature.

Anthropic’s implementation is a version of SynthID-Text, a watermarking method Google DeepMind published in the journal Nature in 2024. It’s part of a broader family of approaches with roots in a 2022 proposal by computer scientist Scott Aaronson.

Critically, the watermark isn’t applied everywhere. Where there’s a single clearly correct answer — a factual detail, a piece of code that would break if a term changed — there’s no room for the “either choice is fine” logic the watermark depends on, so it isn’t invoked. That means factual, terse, or code-heavy text carries a much weaker watermark signal than long-form creative or conversational writing.

Where It Applies

The marking isn’t limited to Claude’s consumer chat interface. According to a support article detailed by Search Engine Journal, it covers output from the Claude API, the Claude apps, Claude Code, Claude Cowork, Claude Tag, and Claude accessed through cloud partners including AWS, Google Cloud, and Microsoft Foundry. Developers building on Claude through the API are just as subject to it as someone typing into claude.ai.

The marking also applies globally, not just to EU-based users. Anthropic says this is because it doesn’t currently have a reliable way to scope the feature by region — so rather than build EU-specific infrastructure, it applied the change everywhere at once.

It’s worth separating this from a related but distinct system Anthropic also introduced: when Claude produces a file, such as a .png, .jpg, or .svg, it attaches a small, cryptographically signed note to the file’s metadata using C2PA, an open industry standard also used by camera makers and photo-editing software. That’s a visible-if-you-look-for-it credential embedded in file metadata — fundamentally different from the text watermark, which lives in the statistical pattern of word choices themselves and survives copying and pasting the words into a new document.

Does It Hurt Writing Quality?

This was the question that drove much of the initial coverage, and it’s worth being precise about what’s actually been shown.

Anthropic’s position is straightforward: in its internal testing, the company says it saw no impact on the content, creativity, or readability of Claude’s text. To back that up, Anthropic points to the research behind SynthID-Text itself. In the Nature paper, Google DeepMind tested the method by rolling it out to a portion of live Gemini traffic — nearly 20 million responses — and comparing user thumbs-up and thumbs-down ratings against an unwatermarked baseline. They found no statistically significant difference. A separate controlled study, with human raters comparing watermarked and unwatermarked answers side by side, also found no detectable quality gap.

That’s real evidence, but it’s evidence about Google’s method as deployed in Gemini — not an independently published, Claude-specific benchmark. Anthropic has not released its own quantitative figures, such as perplexity comparisons or blind preference scores, for its implementation.

Not everyone is convinced the underlying logic holds up regardless of the numbers. Writer and blogger John Gruber argued publicly that the premise is flawed on its face: no two synonyms carry exactly the same meaning, he wrote, so nudging the model to favor “grey” over “overcast” — or vice versa — based on a hidden key rather than the best semantic fit means the system is, by definition, sometimes picking a slightly worse word. That’s a conceptual argument, not a competing data set, but it captures a real tension between “no reader could tell the difference” and “the model wasn’t choosing purely on merit.”

What It Doesn’t Prove

Anthropic is candid about the watermark’s limits, and they matter as much as how it works.

A detected watermark can only establish that Claude was likely involved in producing a piece of text at some point — not who wrote it. It can’t distinguish a passage Claude generated from scratch from one where Claude only lightly edited someone else’s writing. For a journalist who used Claude to tighten a paragraph, or an academic who ran a draft through it for copyediting, a positive detection doesn’t mean the model wrote the substance — but it doesn’t say that on its face either.

The watermark also isn’t durable against everything. Light editing typically won’t strip it out. A full rewrite, where essentially every word is replaced, will — though at that point, Anthropic notes, it’s fair to ask whether the result is still meaningfully “AI-generated” at all. Detectability also weakens as a passage gets shorter, since there are simply fewer word choices for the pattern to live in.

None of this has an opt-out. That’s the detail that generated the sharpest public reaction after the announcement — not the abstract question of whether “grey” beats “overcast,” but the fact that every eligible response, in every private conversation, now carries the mark by default, with no setting to turn it off.

How Claude Compares to Other AI Tools

Anthropic isn’t alone here, and it isn’t first, either.

Google has used SynthID to watermark Gemini’s text output since 2024 — the same underlying method Anthropic has now adapted. Google has also been coordinating with other companies, including Apple, ElevenLabs, Kakao, Nvidia, and OpenAI, to make watermarking tools interoperable across platforms.

OpenAI’s position is less settled. The company’s support documentation, updated around the same time Article 50 took effect, states a goal of extending its “provenance signals” to text models, referencing both Content Credentials and SynthID — but as of this writing, there’s no confirmation that ChatGPT text watermarking has actually shipped. That’s notable given OpenAI’s history: the company built its own ChatGPT text watermarker back in 2024, found it highly accurate, and then shelved it — citing both how easily it could be evaded and concerns that it would disproportionately flag writing from non-native English speakers, whose text tends to show less lexical variety and can trigger some detection methods at higher rates. Whether that same fairness concern applies to Anthropic’s implementation of SynthID-Text hasn’t been independently studied; it’s a documented issue with AI-text detectors in general, not a finding specific to Claude.

Meta, Microsoft, and Mistral all signed the same EU Code of Practice, but none has published rollout details as specific as Anthropic’s or Google’s. xAI has not signed the Code at all.

What’s Still Unclear

Several practical questions remain open. Anthropic has said it will release a detection API so users and third parties can check text for its watermark, but as of this writing, that tool hasn’t launched, and no timeline has been given.

It’s also not clear how the watermark performs across languages other than English. Anthropic confirms that Claude-generated translations do carry the watermark, since every word in a translation is one Claude chose — but the company hasn’t published any comparative data on detection strength or quality impact across different languages, some of which offer far fewer interchangeable word choices for the watermark to work with than English does.

There’s no confirmed minimum length below which detection becomes unreliable, and no indication that a per-customer or per-account opt-out is under consideration.

What This Means Now

For most people using Claude day to day, the watermark is invisible in the most literal sense: nothing about how a response reads, costs, or arrives will change. The available evidence — mainly Google’s large-scale research on the underlying SynthID-Text method — supports Anthropic’s claim that quality isn’t measurably affected, even if that evidence isn’t a bespoke study of Claude itself.

The more consequential shift is structural. AI-generated text is now, by regulatory mandate, expected to carry some form of built-in provenance signal across most of the industry’s major players — and that signal, in Claude’s case as in Google’s, is close to impossible to detect by eye and difficult to fully strip out short of a wholesale rewrite. What’s still missing is a working, publicly available way for ordinary users to check it, evidence about how the system behaves in languages beyond English, and any mechanism for people who’d simply rather not have their private conversations marked at all. Those gaps — not the writing-quality question that dominated the initial coverage — are the ones worth watching as the detection API and cross-language data eventually arrive.

FAQ

Does Claude’s watermark change how it writes or the quality of its answers? Anthropic says no, based on internal testing and on research behind the underlying SynthID-Text method, which found no statistically significant quality difference in a large-scale test on Google’s Gemini. There’s no independently published, Claude-specific benchmark confirming this, and the claim has been publicly disputed on conceptual grounds by at least one prominent critic.

Can I turn off watermarking in Claude? No. As of this writing, there’s no opt-out for individual users or organizations.

Does the watermark apply to the Claude API, or only the chat app? It applies broadly — to the API, the Claude apps, Claude Code, Cowork, Tag, and Claude accessed through AWS, Google Cloud, and Microsoft Foundry. It is not limited to the consumer web or mobile interface.

Can someone tell my document was edited by Claude if I only asked it to proofread? Possibly, but the signal is weaker. Since the watermark only lives in words Claude actually chose, a light copyedit that changes little of the original text leaves less material for the pattern to appear in. Anthropic acknowledges that a positive detection can’t distinguish “Claude wrote this” from “Claude lightly edited this.”

How is Claude’s watermark different from OpenAI’s or Google’s? Google has used the same underlying method, SynthID-Text, in Gemini since 2024. OpenAI has stated an intention to extend similar provenance signals to ChatGPT’s text output but hasn’t confirmed shipping it, and previously shelved its own 2024 text watermarker over evasion and fairness concerns.

Can the watermark be removed or evaded? Light editing generally won’t remove it. A substantial rewrite that replaces most of the original wording will. Independent research on the underlying SynthID-Text method has also identified specific technical attacks that can weaken or defeat detection under adversarial conditions.

Is there a way to check if a piece of text is watermarked by Claude? Not yet publicly. Anthropic has said it will release a detection API, but it hadn’t launched as of this writing, and no release date has been given.