← All articles · Deutsche Version

Your AI Text Is Now Watermarked: What EU AI Act Article 50 Means If You Build Sites, Apps or Content

A practical guide for developers, product teams and content creators. This article is for general information only and is not legal advice.

If your product calls Claude or Grok, the text it gets back is now, or will very soon be, carrying an invisible watermark. You didn't change a line of code. Nothing new appears in the response body. But the words themselves now hold a statistical pattern that the model provider can later detect.

This is not a product whim. It is how the big model providers are meeting Article 50 of the EU AI Act, the transparency rule that has applied since 2 August 2026. For anyone who builds websites, apps or content with generative AI, it raises practical questions. What does the law actually ask of you? What is the provider's job, and what is yours? And what should you check this month?

What changed in the last few weeks

• Anthropic (Claude). Anthropic has signed the EU's Article 50(2) Code of Practice on Transparency of AI-Generated Content. Its Help Center says supported models mark text with an embedded watermark across the API, the Claude apps and Claude Code, worldwide, and that generated files get C2PA Content Credentials where supported. The supported list now includes older models such as Opus 4.7, Opus 4.6 and Sonnet 4.6. Rollout on Amazon Bedrock is set to finish by 10 November.
• ⚠️ Press reports (The Independent and elseif) say the expansion to older models began on 3 November, citing an Anthropic email to customers. I could not find that date on an Anthropic page.
• xAI (Grok). The official API docs say that from 7 November 2026 responses from seven Grok models carry a watermark by default: grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants. It covers plain text, structured outputs and tool-call arguments, on every endpoint.
• OpenAI. ⚠️ The Independent reports that OpenAI has added a feature called "textGrain" to ChatGPT, applied automatically only in Europe. This comes from a single press source; I could not confirm it in OpenAI's own documentation.

Anthropic says about 190 organisations signed the transparency Code of Practice in July 2026. Expect more providers to follow.

How text watermarking works (and where it fails)

A language model writes one token at a time. At many points several next words are almost equally good: "grey" or "overcast". Normally a random number decides. A watermark replaces that random number with one derived from a secret key and the preceding words. Readers see nothing different, but someone holding the key can test whether a passage is consistent with those keyed choices.

Anthropic says its method is a version of Google DeepMind's SynthID-Text. A few points matter for builders:

• Nothing is added to the text. No hidden characters, no markup, no extra tokens and no extra cost, according to both Anthropic and xAI.
• Detection is statistical. xAI's docs say detection returns "watermarked", "not watermarked" or "uncertain". Longer text gives a stronger signal.
• Short text, factual text and code carry little or no mark. If there is only one correct next word, there is nothing to watermark.
• Heavy editing, paraphrasing or translation by another tool weakens it. A full rewrite removes it.
• It is not identifying. Anthropic states the watermark holds no information about the user, organisation or chat.
• Detection is gated. Both Anthropic and xAI grant access to their detectors by application. Anthropic limits it to groups such as regulators, media, researchers and companies that must check their own outputs under the Act.

In short, a watermark is a provenance signal, not proof. A detected mark means a model was probably involved. No mark does not mean a human wrote it.

What Article 50 actually requires

Article 50 sets out four transparency duties. Two fall on providers of AI systems and two on deployers.

1. Article 50(1), providers: systems that interact directly with people, such as chatbots, agents and avatars, must tell people they are talking to AI, unless that is obvious.
2. Article 50(2), providers: generative systems producing synthetic audio, image, video or text must mark outputs in a machine-readable format so they can be detected as AI-generated. The marking must be "effective, interoperable, robust and reliable as far as this is technically feasible". This is the rule behind the watermarks.
3. Article 50(3), deployers: people exposed to emotion recognition or biometric categorisation must be informed.
4. Article 50(4), deployers: deepfakes must be disclosed. AI-generated or manipulated text published to inform the public on matters of public interest must also be disclosed, unless it has gone through human review or editorial control and someone holds editorial responsibility.

The information has to be clear and distinguishable, given at the first interaction or exposure at the latest, and accessible (Article 50(5)).

What does not need a mark

The Commission's FAQ, based on its Article 50 guidelines, lists outputs outside the marking duty: short sequences of numbers, symbols or letters; source code; machine-to-machine outputs never shown to humans; and outputs used only in closed-loop production environments unless they are the final output. Marking also does not apply where the system only performs an assistive function for standard editing or does not substantially alter the input or its meaning.

The dates: 2 August vs 2 December

This is where many summaries go wrong.

• Article 50 has applied since 2 August 2026, for providers and deployers alike.
• 2 December 2026 is not a general deadline. It is a limited transition (Article 111(4), as amended by the AI Omnibus). It covers only generative AI systems already on the market before 2 August 2026, and only the marking and detection duty in Article 50(2).
• Content generated before 2 August 2026 does not need to be labelled retroactively, although the Commission encourages it.

That is why Anthropic and xAI are now retrofitting older models: the grace period for them ends on 2 December.

Fines for breaching transparency obligations can reach €15 million or 3% of worldwide annual turnover, with proportionality for SMEs and small mid-caps.

Provider or deployer: where do you stand?

The split matters, because the watermark is the provider's job, while visible labels are mostly the deployer's.

• A provider develops an AI system, or has it developed, and places it on the EU market or puts it into service under its own name or trademark. This applies even if the provider is outside the EU, as long as the output is used in the EU.
• A deployer uses an AI system under its authority for professional purposes. The Commission's FAQ notes that a freelancer earning regular income counts as professional. Employees of a company deployer, such as web designers, content creators and journalists, are not separate deployers.

So if you publish AI-assisted articles on your site, you are most likely a deployer. Your duty under Article 50(4) is a visible disclosure for public-interest text that nobody reviewed. The provider's machine-readable mark does not satisfy that duty on its own: the Commission says deployers "cannot simply rely" on it for deepfake disclosure.

If you build a product on someone else's API

Here is the hard question. If your app wraps Claude or Grok and you ship it under your own brand, are you the "provider" of a generative AI system that has to ensure marking?

⚠️ This is a question for a lawyer, not for a blog post. The definition of provider turns on placing a system on the market "under its own name or trademark", and the Commission's guidelines add detail I can't apply to your case. Anthropic's own Help Center tells builders to "independently assess what Article 50 requires of your products and services". Treat the model's watermark as a help, not as your compliance.

What this means in practice for your app

Know where your text comes from. List every feature that generates text, images, audio or video: product descriptions, chat replies, email drafts, summaries, alt text, thumbnails.

Log the watermark status. With the Grok API you can read the x-watermark response header. true means the output was watermarked; any other value, or none, means it wasn't. In the Vercel AI SDK it appears in response.headers. Storing it per request gives you evidence later.

Don't opt out casually. xAI lets team admins switch watermarking off per model in the Console. Doing so requires confirming that a legal exception applies and that you are "solely responsible for any marking, disclosure, or other legal requirement". Every change goes into the audit log. Some models may be locked on.

Watch your sampling parameters. Anthropic warns that Sonnet 4.6, and per its limitations list also Opus 4.6, may not reliably carry the watermark when temperature, top_p or top_k is set to a non-default value. If a compliance story depends on the mark, check your configs.

Structured output is marked too. xAI's watermark applies to structured outputs and tool-call arguments. Values are not altered beyond normal token choice, but don't build logic that assumes byte-identical outputs across runs.

Keep C2PA metadata. Claude attaches signed Content Credentials to supported generated files. Image pipelines that re-encode, resize or strip EXIF will often drop it. Preserve metadata if you can.

Add a chatbot notice. If your site has an AI assistant, tell users at the start that they are talking to AI (Article 50(1)), unless it's obvious.

Label what needs labelling. For deepfakes and unreviewed public-interest text, add a visible label. The Commission has published three optional icons for labelling AI-generated content.

What does not need a label

Not every AI touch triggers a disclosure. Under Article 50(4), text needs a label only if it is published, meant to inform the public, and on a matter of public interest, such as politics, public health, consumer safety or economic developments. Even then the duty falls away if it went through real human review or editorial control and a person holds editorial responsibility. The Commission is clear that spell-checking or grammar fixes are not enough: review means examining the substance with relevant knowledge, including fact-checking.

Grammar fixes, translation help and other standard editing assistance are also outside the provider's marking duty. A marketing tagline or a recipe is unlikely to be "public interest". When in doubt, a short transparency note costs little.

Your 11-point Article 50 checklist

1. Inventory every place your product or site generates text, images, audio or video with AI.
2. Identify each model provider and read its marking docs (Claude, Grok, OpenAI and others).
3. Log the x-watermark header (Grok), or the equivalent, per request.
4. Don't opt out of watermarking unless a legal exception is documented.
5. Check sampling parameters (Sonnet 4.6 and Opus 4.6 with non-default temperature, top_p or top_k may drop the mark).
6. Keep C2PA metadata on generated files; don't strip it in image pipelines.
7. Chatbots: tell users they're talking to AI (Article 50(1)).
8. Deepfakes and public-interest AI text: add a visible disclosure, unless the text was human-reviewed under editorial responsibility (Article 50(4)).
9. Generative systems on the market before 2 August 2026: marking by 2 December 2026.
10. Publish a short AI transparency note on your site.
11. Get legal review of your provider or deployer role. ⚠️

Get the free 11-point checklist (PDF)

The bottom line

Watermarks are now on by default for much of the AI text you generate, and that is the provider's work. Your part is different: know which models you use, don't switch the mark off without a reason, tell people when they're talking to a bot, label deepfakes and unreviewed public-interest text, and get a lawyer's view of your role if you ship AI under your own brand.

This article is for general information and is not legal advice. Check the primary sources below, and consult a qualified lawyer about your situation.

Sources

1. AI Act, Regulation (EU) 2024/1689, EUR-Lex: https://eur-lex.europa.eu/eli/reg/2024/1689/oj
2. AI Act Article 50 (unofficial consolidated mirror), artificialintelligenceact.eu: https://artificialintelligenceact.eu/article/50/
3. European Commission, "Transparency obligations under Article 50 of the AI Act" (FAQ): https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
4. European Commission, "Quick Facts: Transparency rules for AI systems": https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems
5. Claude Help Center, "How Claude marks AI-generated content": https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
6. Anthropic, "How Claude's text watermark works" (14 Aug 2026, updated 1 Sep 2026): https://www.anthropic.com/news/claude-text-watermark
7. xAI API docs, "Output Watermarking": https://docs.x.ai/developers/advanced-api-usage/watermarking
8. The Independent, "Anthropic's Claude will now leave a trail that shows whether a text was made with AI" (2 Nov 2026): https://www.independent.co.uk/tech/claude-anthropic-text-watermark-ai-b3077596.html
9. elseif, "Anthropic expands text watermarking to more Claude models": https://www.elseif.net/anthropic-expands-text-watermarking-to-more-claude-models