EU AI Act and Claude Watermarking: How Translation Watermarks Work And Why It's Important For Your Translations

Aug 20

7 min read

Claude now embeds an invisible, machine‑readable watermark in the text it generates. This isn’t a stylistic choice; it’s a compliance move driven by the EU AI Act, which requires providers of generative AI systems to make their outputs detectable as AI‑generated.

Anthropic has stated that all Claude models launched on or after 2 August 2026 support this marking at launch, and that older models are being retrofitted. The watermark is applied globally, not just for EU users, and travels with the text when it’s copied, pasted, integrated or even completely retyped into other tools.

The Background: EU AI Act and Transparency Rules

The immediate trigger is Article 50 of the EU AI Act, which took effect for new systems on 2 August 2026. It obliges providers of AI systems that generate synthetic text (and audio, image, or video) to ensure outputs are “marked in a machine‑readable format and detectable” as artificially generated or manipulated.

To operationalize this, major AI providers—including Anthropic—signed the EU Code of Practice on Transparency of AI‑Generated Content in July 2026. Non‑compliance can lead to significant fines (up to €15 million or 3% of global annual turnover), which is why companies are rolling out watermarking across their entire product surface, not just in Europe.

How Claude watermarking works at a high level

Claude’s watermark is not a visible tag or a hidden comment like the name suggests. Instead, it’s a statistical pattern embedded in the model’s word choices.

Invisible to readers: Anthropic says the watermark does not change the meaning, quality, or readability of the output.
Model‑level embedding: The pattern is created during generation by biasing token selection in a way that’s undetectable to humans but detectable with the right key and algorithm.
Persists through copy/paste: Because the watermark is the text itself, it travels when you copy and paste, and may persist through editing.
Detection requires a key: Third‑party detection (once available) will rely on a secret key and statistical tests (e.g., z‑scores) to estimate the likelihood that Claude was involved.

Anthropic is using an approach aligned with Google DeepMind’s SynthID Text research and plans to release a detection API so platforms and users can check for Claude’s mark.

Too complicated? Here is the simplified version.

When a model writes text, it does not pick one fixed next word every time but ranks many possible next tokens, then chooses from them. A token can be a word, part of a word, punctuation, or another small unit of text. For example, after the phrase “The weather today is,” the model might consider words like “sunny” “cold” “clear” or “mild”.

A watermarking system can gently guide and influence those choices. It may favor one set of acceptable tokens over another in a pattern known only to the watermarking system. The output still reads naturally, but the distribution of words carries a hidden signal.

A detector later checks whether the text contains this signal. It does not need to know what prompt created the text, because it looks for statistical patterns that are unlikely to appear by chance in ordinary human writing.

Think of it like a very subtle rhythm in the text. A reader does not hear it, but a measuring tool can.

This kind of watermark has to balance several goals:

It should not make the writing worse.
It should work across many topics.
It should be hard to remove by accident.
It should avoid false accusations.
It should still work when the output is not very long.

That last point is hard. Short text gives a detector less evidence. A one-sentence answer may not contain enough signal to judge, whereas longer passages usually give detection tools more to work with. This is precisely why the EU demands the watermark to work on a paragraph level.

What This Means For Other LLMs And Translation Tools

We expect a wave of detection tools and platform integrations in the coming months:

Detection APIs: Anthropic has said it’s working to enable users and third parties to detect Claude’s embedded watermarks.
Platform policies: Marketplaces, publishers, and content platforms may start requiring or encouraging AI‑content labeling, using these signals to enforce policies.
Other LLMs will follow: To avoid regulatory risk and potential fines, other frontier model providers are under pressure to implement similar marking schemes.

In practice, this means AI‑generated text will become increasingly “identifiable” at the ecosystem level, even if individual detectors aren’t perfect.

And YES, this applies to translations, too!

This is crucial for authors: AI‑translated text is watermarked when produced by a supported large language model.

Translations may look like a simple conversion, but AI translation involves the use of generative AI, not just assistive. The model reads the source, interprets meaning, chooses phrasing in the new language, and produces a fresh target-language text.

That also includes services and workflows that rely on Claude (directly or via API) for translation, such as tools many authors use for a quick translation turnaround. If a service like ScribeShadow (or similar AI translation/editing tools) uses new Claude or other watermarking LL models under the hood, its output will carry the watermark

Because the watermark is embedded at the paragraph/text level via token choices, it’s not something a standard proofread or light edit will reliably remove. As of right now, it is unclear if severe editing is going to be enough to get rid of the watermark.

Important: This also means the watermark is not copied from the source text, but embedded into the translated output itself. If the English original has no watermark, the AI-generated German translation will have one. If the English original was already AI-generated, the translation will receive a new watermark signal in the target language.

Can a Watermark Be Edited Out?

The short answer: not easily, and not reliably.

The watermark is woven into the statistical structure of the text, not stored as metadata you can strip.
Anthropic notes it “may persist through editing,” and only heavy rewriting is likely to degrade or break the signal.
Detection works at the passage level, so small tweaks won’t reliably “clean” the text.

For authors, this means: if your book or manuscript was translated with AI, that signal may remain detectable even after a standard proofread. If you want to keep the use of AI unknown to readers (for fear of rejection), it will become more and more important to find an actual editor who knows what they're doing. And even then, it may not be enough.

What Authors Should Do Now With Claude Watermarking Translations

If you’ve used AI to translate your text, it’s time to rethink your workflow.

1. Decide on your transparency stance

Ask yourself:

Do I want my book to be clearly human‑translated/human‑edited?
Am I comfortable with the possibility that AI involvement could be detectable?

Your answer will guide how aggressively you need to rework the text.

2. If you used AI translation before, consider a full human rewrite

If your current translation was produced by AI:

Option A: Re‑translate with a human translator A professional translator who works from your source text and rewrites in the target language will produce text without an AI watermark. This is the cleanest solution, albeit the most expensive one.
Option B: Commission a deep, creative edit (not just proofreading) Hire an editor who:
- Rewrites sentences and paragraphs, not just fixes grammar.
- Rethinks phrasing, rhythm, and style in the target language.
- Treats the text as a new creative draft, not a light polish.

A “native proofreader” who mostly corrects typos and grammar is unlikely to remove the watermark signal, because they’re not changing the underlying token patterns enough. Beware of editing companies, too. Most countries, like Germany, do not issue certificates for the job as an editor, so anyone can call themselves that. A thorough edit of an AI translated 70k - 80k novel will provide over 10,000 tracked changes minimum.

3. Move important projects off AI translation before new models roll out

If you’re mid‑project and still using AI translation:

Finish and lock in your current AI‑based projects now, before newer, more strictly watermarked models become the default.
Then invest in a human translator or deep editor to produce the final, publishable version.

This gives you a clear before/after line: early drafts may be AI‑generated; the published text is human‑crafted.

4. Update your contracts and briefs

When working with translators or editors:

Specify that you want original human translation/editing, not AI‑polished output.
Ask whether they use AI tools and, if so, how (e.g., terminology lookup vs. full‑text generation).
For sensitive projects, request a statement that the final text is not AI‑generated.

This isn’t about punishing AI use; it’s about transparency and trust.

Readers, retailers, and regulators increasingly want to know when content is AI‑generated.
Watermarking makes it easier to distinguish between human‑crafted and AI‑generated work at scale.
For authors, this shifts the value back toward human translators and editors who can genuinely rewrite, reshape, and localize your voice.

If your brand as an author rests on authentic, human storytelling, the smart move is to treat AI translation as a drafting aid at best—and to invest in human professionals for the final, publishable text.

Practical Next Steps For Authors Using AI To Translate

For most authors, the right response is not panic. This changes the way authors use AI to translate, yes, but it also means that the book world is shifting back to wanting human generated output (or, the very least, the transparency if it's not).

Here's what you should do:

Audit your current translations and drafts: If you want to use a model without watermarking, now is the time. Most models will likely implement watermarking by Dec 2nd.
For books you plan to publish or republish, budget for human translation or deep editing.
Talk to your editor/translator about AI use and agree on a human‑first workflow.
If you’re still experimenting with AI translation, finalize those drafts soon, then move to human rewrites for the final version.

The goal isn’t to ban AI from your process, but to ensure that what reaches readers carries the voice and craft you intend—and that you’re in control of how “AI‑generated” your published work really is. And most of all, be honest with yourself and treat watermarked translations as what they are: AI-generated language outputs that still deserve human responsibility.

If you need any help finding a good editor or translator, the entire Literary Queens team is always here for you.