Claude · The deadline tracker

EU AI Act text watermarking: what writers need to know

The law is real, the dates are fixed, and one delay changes things for existing systems.

Article 50 of the EU AI Act sets the transparency rules for AI-generated content. They take effect August 2, 2026. A voluntary Code of Practice puts them into practice, and Anthropic signed it. The final version of the Code came out June 10, 2026. Roughly 190 companies and organisations had signed by the end of July 2026. The Code asks providers to mark machine-made output. The marks need to be machine-readable and detectable as artificially generated, where that is technically possible. Claude models launched in the EU on or after August 2, 2026 carry machine-readable marks at launch. Text gets a watermark. Files get C2PA metadata. Older models are still being updated. One timing point matters for writers. The EU 'AI Omnibus' simplification pushes watermarking back for systems placed on the market before August 2, 2026. Those systems get until December 2, 2026. New systems comply from August 2, 2026. For writers, the practical effect stays the same. The marks are real, they have documented limits, and editing remains the honest approach.

Why this page exists

Searchers are looking around the August 2, 2026 enforcement date. They want the dates and what those dates mean for their writing, stated plainly. The search mix includes legal confusion and vendor announcements.

The tells that show up here

Not every AI tell appears everywhere. These are the ones that cluster on this surface, and what it is about the format that produces them.

vague-attribution
'New EU rules say' without naming the provision and the date.
not-only-but
The parallel tic when distinguishing the Act from the Code of Practice.
throat-clearing
News-style openers that delay the date.

Before and after

A generic AI draft on the left. On the right, the real output of running that draft through underslop, checked so it adds no name or number the draft did not already carry. The voice is The Steady Hand.

Before · generic AI draft

Are you confused about the new EU watermark rules? Don't worry — you're absolutely right to ask! In today's rapidly evolving regulatory landscape, it's not just about compliance; it's about staying ahead of the curve! Experts agree that understanding Article 50 is the cornerstone of navigating AI's brave new future!

After · edited

The new EU watermark rules can be hard to follow. Article 50 is the key part, and it's worth reading it carefully. Plain compliance matters, and staying current with the rules keeps things simple. Most guidance points there first, so starting with that article is a practical move.

How to do it by hand

  1. State the effective dates The transparency obligations fall due on August 2, 2026. The watermarking grace for systems already on the market runs to December 2, 2026, and it covers C36 and C37. The final Code was published on June 10, 2026, under C38.
  2. Name what the law requires The Code requires machine-readable marking that can be detected as artificially generated, where technically feasible.
  3. Connect it to Claude Note that Anthropic signed the Code. EU-launched models from August 2, 2026 will mark output at launch, and older models will be updated.
  4. Give writers the practical takeaway Track which of your tools leave marks on output, and keep your workflow open about it. Marks fade with heavy editing, paraphrasing, translation, and format conversion. Editing stays the honest choice, and that's the rule to keep.

Questions

What does the EU AI Act say about watermarks?

Article 50's transparency rules take effect on August 2, 2026. The Code of Practice asks providers to label machine-made output. The label must be readable and machine-detectable. That applies where the technology allows it.

When did the EU watermark rules take effect?

The transparency obligations start August 2, 2026. The AI Omnibus simplification pushes watermarking back to December 2, 2026 for systems already on the market before that date. New systems follow the rules from August 2, 2026.

Does the EU AI Act apply to Claude?

Anthropic signed the Article 50(2) Code of Practice. Claude models launched in the EU on or after August 2, 2026 carry machine-readable marks at launch. Older models are being updated.

Does the law apply to me as a writer?

Marking duties sit with providers, not with the people using their tools. Deployers do label in certain cases, such as deepfakes and public-interest text. For most writers, the practical change is small. The Claude tools you use now mark output. The marks have known limits.

Who signed the Code of Practice?

About 190 companies and organisations had signed by the end of July 2026. Anthropic was among them. The Code's final version came out on June 10, 2026.

Claude marks are real and the coverage is still settling. underslop is the edit, not a remover: it edits an AI draft so it reads like you wrote it, and it makes no promise about what any detector will report. Your text is held for 0 seconds.

Does Claude watermark text? Yes, and here is what that means