# EU AI Act text marking and Article 50

Originally published at https://gifi.ai/guides/eu-ai-act-text-marking.

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](https://gifi.ai/authors/can-balkaya). Published 2026-08-17; updated 2026-08-17.

## How to verify the job

1. **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.
2. **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.
3. **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.

## 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

### 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.

## Primary sources

- [EU Code of Practice signatories](https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content)
- [Anthropic on Claude text watermarking](https://www.anthropic.com/news/claude-text-watermark)
- [Google DeepMind on SynthID](https://deepmind.google/models/synthid/)