Required by EU Law—and Fines Start August 2

Why the sudden change? Article 50 of the EU AI Act, which covers "transparency obligations," is the answer. Specifically, Article 50(2) sets a clear requirement for model providers: outputs from generative AI systems must carry "effective, reliable, robust, and interoperable" machine-readable markers so that AI-generated or manipulated content can be identified.

This obligation becomes enforceable from August 2, 2026, overseen by market surveillance authorities in each member state. That's exactly where Anthropic's "August 2, 2026, and later" line comes from.

Signing the Code is voluntary, but Article 50 obligations are mandatory. Signing has a direct benefit: the Commission and the AI Office have already deemed the Code "appropriate," letting signatories use it to demonstrate compliance. Non-signatories must prove their alternative measures suffice, case by case, to national regulators.

So viewing Claude's watermark as just "Anthropic added a new feature" misses the point. This is a whole industry moving in step, pushed by the same law at the same time. Anthropic just happens to have documented its approach publicly.

Two Types of Markers: Watermarks in Text, Signatures on Files

Since it's required by law, what exactly is Anthropic doing? Two technologies, each handling a different medium.

This distinction determines how much each marker can survive. Try the demo below to see what happens to each under common scenarios.

How an Invisible Watermark Gets Embedded in Text

An invisible mark that doesn't change the meaning sounds contradictory: if the words are the same, where can the mark hide?

The key insight: when a model writes a word, it doesn't have just one choice. It calculates probabilities for a set of candidate words and picks one. Ask the model to write the same sentence twice, and you might get different word choices. This "selection freedom" is where the watermark hides.

The method works by subtly biasing this choice: a key, combined with the last few words written, generates a random number. This number scores each candidate word, and the higher-scoring candidate wins the selection process. The chosen words are still perfectly reasonable, the sentence still reads naturally, but collectively, the word choices carry a statistical bias.

With this principle in mind, the official list of limitations makes much more sense. They aren't just dry disclaimers—they're direct consequences of how the mechanism works.

Finding the Mark Doesn't Mean AI Wrote It

So can this be used to determine if a piece of writing is AI-generated? Here's the official statement: detecting the marker means the content may have been processed by Claude. It does not, by itself, confirm the complete origin of the content.

Two situations distort the signal. First, Claude might not have been involved in the original creation: people routinely use it to proofread, translate, summarize, or reformat, so text and ideas that originated elsewhere still carry Claude's marker on the output. Second, content might have been edited, excerpted, or mixed with other material after Claude processed it.

You can click through these scenarios to see when the marker is present and how to interpret it.

The Absence of the Mark Doesn't Mean It Wasn't AI

The reverse is also false. The official doc lists five cases where content was genuinely generated by Claude but won't have a detectable marker:

The fifth case is platform differences: content generated on a platform, feature, or file type that doesn't support that marker type will also lack the marker. The official doc applies the same caveat to cloud platforms: when accessing Claude via AWS, Google Cloud, or Microsoft Foundry, the text watermark is applied, but the signed provenance metadata "may not be supported on all platforms."

So the Watermark is a Source Signal, Not a Plagiarism Detector

Since you can't infer reliably in either direction, how should you use it? This mechanism provides a source signal, not an AI determination.

These limitations come straight from Anthropic's own document. They're not external criticisms. And they were written with unusual candor. One section is even titled "Limitations," with the cases listed in both directions, including the admission that "detecting a marker doesn't confirm the full origin of content."

"Detection of a Claude marker indicates the content may have been processed by Claude. It does not, by itself, confirm the complete origin of the content."

Claude Help Center, "How Claude marks AI-generated content"

Looking at the bigger picture, this reflects the design intent of the EU law: Article 50 demands detectability, not condemnation. It adds a layer of source information to the information ecosystem, giving people an extra data point when consuming content—not a hammer for judgment.

If You Build Products on Claude, You Have a Duty Too

One final note for those integrating Claude into their own products. The document is unambiguous here: if you deploy Claude into your product, you are responsible for evaluating what Article 50 requires of your product and service.

This reflects the liability structure under EU law. It divides the generative AI chain into two ends with different obligations:

Per its commitments under the Code, Anthropic aims to support you in fulfilling your own transparency obligations, with technical guidance to be released gradually. Until then, a watermarked model does not equal a compliant product.

EU law mandates it—and Anthropic isn't alone

This comes from Article 50 of the EU AI Act, which requires model providers to add machine-readable markers to generated content. It's enforceable from August 2, 2026. Signing the accompanying transparency Code of Practice is voluntary, but the Article 50 obligation is mandatory. Signatories can use the Code to demonstrate compliance.

This is an entire industry moving in step under one law, not a single company’s choice.

How a sentence gets subtly marked

There are two separate layers. The text watermark is part of the text itself and survives copy-paste. File signatures use the C2PA standard and are attached to files like .svg, .png, and .jpg. They can be stripped by a format conversion or a screenshot. Anthropic hasn't disclosed its algorithm. To illustrate the general principle, we use a similar public technology (Google SynthID-Text, Nature 2024). This is not Anthropic's implementation.

At each step, the model picks a word from several candidates. The watermark uses a secret key to score these candidates, letting higher-scoring words win—while the sentence stays completely natural. Illustrative figure. Based on the public paper for Google SynthID-Text (Nature, 2024), not Anthropic's implementation; candidates and scores are illustrative.

Finding it isn't proof. Not finding it isn't proof.

Anthropic states this directly in the doc: detecting the marker only indicates content may have been processed by Claude.

If you build on Claude, you have obligations too

The marker at the model layer is Anthropic's duty. But if you deploy Claude in your product, you might fall under Art. 50(4) (deployer obligations). Deepfakes, and AI-generated text on public interest topics published without human review, require your own labeling.