EU AI Act · Article 50 · 2 December 2026
MarkdownThe duty is on the provider. A rewrite does not meet it.
Article 50 of the EU AI Act asks providers of generative AI systems to mark output in a machine-readable way so it can be detected as artificially generated. The Code of Practice is one way to show that. The duty sits with the provider, not with a later editor. From 2 December 2026 the marking rule for this class of system is in force. Gifi can change words or strip a local file note. That does not satisfy a provider’s marking duty, and it does not remove a disclosure a platform or a law still requires.
Reviewed by Can Balkaya · Published
Process
How to verify the job
01
Decide who you are
If you sell a generative system in the EU, Article 50 is your problem. If you are editing text you own, you are not discharging that duty by running Gifi.
02
See what the provider already marked
Claude and Gemini watermark text. ChatGPT marks images and supported audio, not the words. File credentials are a different surface.
03
Keep the disclosure if you still owe one
A rewrite or a metadata strip does not retire a label the law, a contract, or a platform asked you to keep.
Scope
Supported and unsupported surfaces
| Surface | Gifi action | Verification limit |
|---|---|---|
| Named metadata or invisible Unicode | Inspect, remove, and report named actions | Reinspect the returned bytes independently |
| Statistical patterns in ordinary words | Rewrite the prose | No certificate against a private detector |
| Visible logos, pixels, audio, or video | Unsupported | No removal claim |
What Article 50 actually asks
The Act does not name “SynthID.” It asks for a machine-readable mark that is effective, interoperable, robust and reliable enough to detect synthetic output. Watermarking is how the Code of Practice proposes to meet that for free-form text, which has nowhere to put a metadata packet.
The Code also talks about a length floor. A two-line reply may carry no reliable text mark. A long draft is the case the rule is written for. Treat short chat as a weak signal, not as proof of anything.
Who already marks text
Google has applied SynthID-Text in Gemini. Anthropic watermarks text from Claude models launched on or after 2 August 2026, at the model layer, with no opt-out. OpenAI has not shipped a text watermark. Microsoft, Meta and Mistral signed the Code and have not published a text method.
That spread is why a single “AI watermark remover” page that does not name the lab is usually selling confusion.
What Gifi will not claim
Using Gifi does not make you compliant. It does not mark output for you. It does not un-mark output in a way a regulator would accept as “the content was never AI.” A rewrite changes words. A strip deletes named bytes. Those are technical facts.
If you owe a disclosure, keep it. If you do not own the content, do not process it here.
Questions this usually raises
- Does running Gifi satisfy Article 50?
- No. The duty is on the provider of the generative system. An editor’s rewrite is not that mark.
- What happens on 2 December 2026?
- The marking rule for this class of system applies. Providers who serve the EU need a machine-readable mark. Users of Gifi do not inherit that duty, and they do not discharge it either.
- If I rewrite Gemini text, is the legal mark gone?
- The words changed. That is all we can say. We cannot certify Google’s detector, and we cannot retire a disclosure you still owe.