Around 190 signatories put their names to the EU Code of Practice on Transparency of AI-Generated Content in July 2026. On August 2, the marking obligation became live for providers serving the European market. Two weeks later, on August 14, Anthropic published a detailed explainer on how it is complying — and the answer is not a disclaimer, a metadata tag, or a visible label. It is a pattern woven into the word choices themselves, invisible to every reader, and detectable only by whoever holds the key.

The Rule That Quietly Went Live on August 2

This Is Not a Claude-Only Story

The single most misread part of this announcement is that it is an Anthropic policy choice. It is not. Anthropic states plainly that it is implementing watermarking to comply with the EU AI Act, and that the requirement lands on every AI system provider serving the EU market. Other major model developers signed the same Code of Practice and will roll out their own watermarks, each with a different key and potentially a different technique entirely.

Anthropic is also applying the watermark globally rather than only inside the EU, and it is candid about why: it does not yet have a durable way to scope the behavior by region. So a developer in Seoul, Sao Paulo, or San Francisco gets the same watermarked output as one in Berlin. The company says it will keep evaluating approaches and share updates.

Older Models Get a Grace Period

The law includes a transition period for models launched before August 2, 2026. Anthropic says it is working to add watermarking to those earlier models as well, rolled out over the coming months. In practice, that means the installed base of production systems running on pre-August models will shift into scope gradually rather than overnight — a detail worth noting if you are auditing an existing pipeline.

How the Watermark Actually Works

Same Randomness, Different Source

A language model picks one word at a time from a list of plausible candidates. Anthropic’s example: in the sentence “The weather today was cold and…”, the next word is very unlikely to be “sugary” but quite likely to be “overcast” or “grey.” Between those two, the meaning barely changes, and the choice is normally settled by a random number.

Watermarking changes the source of that randomness. Instead of an arbitrary random number generator, the model uses a secret key plus the few words that came before to settle the pick. The output still looks random. But anyone holding the key can check whether a passage’s sequence of choices is consistent with what Claude would have produced under that key, and assign a probability that Claude was involved.

Anthropic offers a board-game analogy. Imagine playing Monopoly where, instead of rolling dice, players read successive digits of pi starting from a randomly chosen position. To the players, the moves are indistinguishable from real rolls. But an observer who knows pi could later inspect the full sequence and conclude the game was almost certainly driven by pi. The game is, in effect, watermarked.

Where the Watermark Cannot Live

The technique only works where a genuine choice exists. Anthropic gives the example of “Isaac Newton’s most famous work was called Principia…” — the next word must be “Mathematica,” so there is nothing for the watermark to act on. The same applies to arithmetic: once the model writes “2 + 2 =”, there is no equally good alternative to “4,” and the nudge is simply not applied.

This has a direct consequence for developers. Code, which usually has to be exact, carries generally less watermarking than prose. Where arbitrary naming choices do exist — comments inside code, for instance — the watermark can attach, but Anthropic says the effect on the actual code produced is negligible. Dense factual passages are similarly sparse.

The Quality Question, Answered With Data

The method is a version of SynthID-Text, introduced by Google DeepMind in a Nature paper in 2024, and belongs to a family of approaches tracing back to a 2022 proposal from Scott Aaronson. In the SynthID-Text work, DeepMind served a watermarked model to a portion of Gemini traffic and compared thumbs-up and thumbs-down ratings against the unwatermarked baseline. They found no statistically significant difference. A controlled study with human raters comparing watermarked and unwatermarked answers side by side also found no quality gap. Anthropic reports the same in internal testing: no impact on content, creativity, or readability.

What This Changes If You Ship AI Content

Detection Is Promised, Not Shipped

Anthropic says it will “soon” offer a watermark detection API and is still working out implementation details. Until that exists, the practical situation is asymmetric: the mark is going into outputs now, while the ability to check it sits with whoever holds the key. Detection also degrades on short samples — fewer word choices means less signal — and confidence rises as passage length grows.

Equally important is what a detected watermark does not prove. Anthropic is explicit that it can only answer “what is the likelihood this was partly written by Claude?” It cannot confirm that text was human-written, cannot identify a different AI, and cannot distinguish “Claude wrote this” from “Claude heavily edited this.” Ownership and legal responsibility are unchanged.

Editing, Translation, and Files

If Claude only proofreads your grammar and punctuation, nearly all the words remain yours and there may be too little for a watermark to attach to. Translation is the opposite case: every word is chosen by the model, so translated output carries a full watermark. Light editing probably will not strip the mark; a complete rewrite replacing every word will — at which point, as Anthropic notes, it is arguable whether the text is still AI-generated at all.

Files follow a different mechanism entirely. When Claude produces a supported file type such as .png, .jpg, or .svg, it attaches a C2PA content credential — a cryptographically signed note in the file metadata recording that the file was made or processed with Claude. Nothing in the pixels changes, and any C2PA-aware tool can read it. That is a label, not a watermark, and the distinction matters when you are writing a compliance policy.

The Reaction Nobody Priced In

Anthropic first said it would watermark text on August 11. By August 12, TechCrunch was reporting that some Claude users were angry the marks could expose their AI use at work or in class — which is, of course, precisely the transparency outcome the EU rule was written to produce. The friction is not technical. It is that a large population had been operating on the assumption that AI authorship was untraceable, and that assumption has an expiry date now.

Worth noting alongside this: on August 14, TechCrunch reported that Google will let users remove the visible watermark from its AI generations. Visible marks and statistical or cryptographic ones are moving in opposite directions — the ones you can see are becoming optional, while the ones you cannot see are becoming mandatory.

Related

• 60 AI Agents Just Attacked Math’s 150-Year Mystery
• Meta Just Made the Cloud Optional for AI Agents
• HappyRobot Raises 150M for Enterprise AI Agents
• Tech Digest — August 13, 2026
• Tech Digest — August 12, 2026

Sources

1. Anthropic — How Claude’s text watermark works (Aug 14, 2026)
2. TechCrunch — Anthropic shares more details about how Claude’s new watermarks will work (Aug 15, 2026)
3. TechCrunch — Some Claude users are mad that Anthropic’s new watermarks will catch them (Aug 12, 2026)
4. TechCrunch — Google will now allow users to remove visible watermark from its AI generations (Aug 14, 2026)
5. Nature — Scalable watermarking for identifying large language model outputs (SynthID-Text, 2024)