EU AI Act Article 50: why every provider is watermarking text now
The regulation that became enforceable on 2 August 2026, what it requires, and why Anthropic applied it worldwide instead of only in Europe.
Anthropic's watermark · 2 min read
If you are wondering why text watermarking appeared across the industry in August 2026 rather than gradually, the answer is a deadline.
What Article 50 requires
EU AI Act Article 50 covers transparency obligations for generative AI. The core requirement: providers of generative AI systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.
It became enforceable on 2 August 2026 for newly launched systems. That date is why Anthropic's cutoff is models launched on or after 2 August 2026.
Two things worth understanding about the drafting:
"Machine-readable" is the operative phrase. The obligation is not a visible label. It is a signal a machine can find — which for text means something embedded in the output itself.
It applies to providers, not users. The obligation lands on whoever supplies the model. You are not breaking the regulation by editing text; the provider satisfies it by marking output.
Why global rather than EU-only
Anthropic applies the mark to all users worldwide, not just European ones. This is a choice, not a requirement, and there are straightforward reasons for it:
- Routing is unreliable. Deciding per request whether a user falls under EU jurisdiction is fragile — VPNs, travel, corporate proxies, API calls from anywhere.
- Model-level is simpler than surface-level. Anthropic marks at the model, so every product built on it inherits the behaviour automatically. Doing it selectively would mean maintaining two behaviours.
- Other jurisdictions are moving the same direction. Building one compliant behaviour is cheaper than building several.
The practical result: the EU regulation set the standard for everyone.
What it does not require
Worth being precise, because a lot of commentary overstates it:
- It does not require providers to publish the detection method.
- It does not require the mark to be robust against editing. Article 50 asks for marking; it does not specify durability thresholds.
- It does not make editing AI output illegal.
- It does not require users to disclose AI assistance. Separate rules may, and institutional policies certainly may, but that is not this article.
That first point explains the situation users are in: a mark exists, and no public tool can check for it.
What this means practically
Text generated by a current Claude model carries a mark. Anthropic states it may not survive heavy editing, paraphrasing, translation, mixing with other content, or very short passages.
Your own institutional rules are a separate question from the regulation. If your university, employer or publisher requires disclosure of AI assistance, that requirement stands regardless of what any watermark does or does not survive — and satisfying a disclosure rule is not something a text tool can do for you.
Related: does Anthropic watermark Claude's text? · what SynthID does
Sources: Anthropic Help Center · TechCrunch, 11 Aug 2026