EU AI Act watermarking rules: Article 50 guide
Updated August 20, 2026 · 9 minute read
Article 50 of the EU AI Act introduces transparency duties for certain AI systems and AI-generated or manipulated content. The rules are broader than adding one invisible character, and they do not create a public promise that every AI passage can be identified by any generic website.
This page is educational, not legal advice. Use the European Commission's current guidance and qualified counsel for compliance decisions.
When the rules apply
The European Commission's Article 50 questions and answers says the transparency obligations apply from August 2, 2026. A limited transition until December 2, 2026 applies to marking and detection duties for systems placed on the market before August 2.
The Commission's quick facts page groups duties around informing people when they interact with AI, machine-readable marking of generated or manipulated content, and specific visible disclosures by deployers.
Provider marking and deployer disclosure are different
Under the Commission's explanation, providers of systems generating synthetic audio, image, video, or text must design outputs to be marked in a machine-readable format and detectable as AI-generated or manipulated, as far as the legal requirements apply.
Deployers have separate visible disclosure duties in defined situations. Examples include deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest without human review or editorial control.
A machine-readable mark alone may not satisfy a visible disclosure duty. Role, purpose, editing, audience, and exceptions matter.
What “machine-readable” can mean
The Commission's Code of Practice on transparency of AI-generated content addresses marking and detection across audio, image, video, and text. Technical approaches can include metadata, content credentials, robust watermarks, or other interoperable signals.
Different media need different methods. C2PA uses signed manifests for provenance. Statistical text watermarking can alter token probabilities. Neither should be reduced to “delete every invisible character.”
What Scan AI Text does
Scan AI Text checks inspectable artifacts in supplied text and supported files. It can report hidden Unicode, unusual spaces, structured metadata, and file metadata fields. It can clean supported items after showing a before-and-after preview.
It cannot certify Article 50 compliance, verify every vendor's machine-readable mark, or prove that a passage is human-written. Cleaning content may remove useful provenance or conflict with disclosure duties. Preserve originals and required labels.
Practical checklist for publishers
- Identify whether you are a provider, deployer, or ordinary personal user.
- Map which content types your system generates or manipulates.
- Preserve provider-supplied provenance and machine-readable marks.
- Add visible disclosure where your use requires it.
- Record human review and editorial control.
- Test accessibility and placement of labels.
- Monitor Commission guidance and national enforcement.
- Get legal advice for your actual business and jurisdiction.
Frequently asked questions
Does Article 50 mean every AI-written email needs a watermark?
No. Scope, exceptions, role, and use matter. The Commission guidance excludes some outputs and explains when provider and deployer duties apply.
Can Scan AI Text tell me whether a document complies?
No. It can produce technical evidence about supported artifacts, not a legal compliance determination.
Should I remove an AI watermark before publishing?
Do not remove provenance or disclosure required by law, contract, platform rules, or professional standards. Use inspection to understand content, not to conceal origin.