What Article 50 asks for
Article 50 of the EU AI Act deals with transparency and labelling. It requires providers of generative AI systems to ensure that their outputs are marked in a machine-readable way, where applicable. It also places disclosure duties on deployers in specific situations, including when they publish AI-generated or manipulated content to inform the public on matters of public interest. Those duties are not a blanket requirement to label every synthetic image, recording, video or piece of writing published in Europe.
The rules also contain exceptions and qualifications. The EU/OpenAI provenance guidance does not require watermarking for outputs shorter than 200 tokens, roughly 150 English words, or for code snippets. Text that has undergone human review and editorial control can also fall under the relevant exception where a person or organisation holds editorial responsibility. The exact duty therefore depends on the system, the content and how it is published.
The European Commission's Code of Practice on Transparency of AI-generated Content is intended to help providers and deployers apply those requirements. It should not be read as a rule that every teacher showing a ChatGPT-written summary, or every studio using generated images in a client deck, automatically has to attach the same label. The circumstances and the applicable exception matter.
How textGrain is supposed to work
OpenAI's announced method is called textGrain. Rather than adding a visible notice to the end of an answer, it is intended to change small decisions made while the text is produced, so the wording itself carries a pattern a detector can read. According to Tubefilter's report, OpenAI says the mark should survive copy-pasting and reformatting. That remains a company claim: no independent testing has established the copy-paste durability. OpenAI also says the method does not expose information about the person who asked.
That second point carries weight in Europe. The GDPR sets the baseline for what a company may store about a user, and a watermark that identified a specific person would sit badly beside data-minimisation rules. Marking the text rather than the user avoids that particular concern, although it does not remove the need to handle provenance data carefully.
Video got labels first
YouTube and TikTok already attach labels and provenance data to AI-generated clips, mostly through the C2PA standard. The EU's position is that written output should not get an easier ride than a video, but Article 50 does not make the obligations identical in every case. The comparison runs through the Tubefilter report on OpenAI's announced commitment, not a statement that a fully deployed text-watermarking system is already operating across ChatGPT.
Fines and enforcement
Under Article 99(4) of the EU AI Act, breaches of obligations other than the prohibited-practice rules can attract an administrative fine of up to €15 million or 3% of a company's total worldwide annual turnover, whichever is higher. That is a maximum, not an automatic charge for every Article 50 breach. The competent authority must apply the rules proportionately and take the circumstances into account. The provision can apply to relevant providers or deployers, depending on which obligation was breached; it does not simply place every duty on the company offering the model.
For a small European studio, that figure is theoretical. Its practical force is aimed at organisations that provide or deploy systems at scale. The European AI Office holds direct powers over general-purpose model providers, while national authorities remain important for other AI Act obligations and actors.
The December target
OpenAI has announced a target of 2 December 2026 for its planned text-provenance implementation. That is a date to watch if you use ChatGPT in a workflow that produces published material, but it is not evidence that the system is already fully deployed. Nor does the date turn every ChatGPT answer into a legally label-required publication: the Article 50 exceptions and the distinction between providers and deployers still apply.
Text from a model running on your own machine does not pass through ChatGPT, so OpenAI's planned watermark would not be in it. Locally installed models are what we exercise in our AI Arena benchmark rig, and textGrain does not change how they behave.
What remains unsettled
The transparency rules are separate from the wider timetable for high-risk AI systems. Reported changes to that timetable should not be presented as settled Article 50 deadlines. For this article, the important point is narrower: the transparency obligations apply to the relevant providers and deployers, while OpenAI's text-marking system remains an announced implementation rather than a fully verified, universally deployed feature.
No independent test of textGrain has been published. Watermark detection for text is less developed than it is for images, where C2PA metadata can be read by anyone with the right tool. The claim that the mark survives copy-paste comes from OpenAI and the reporting around it, not from a third-party benchmark.
Does the watermark apply to writing I do myself?
No. Article 50 concerns content generated or manipulated by AI systems, not text a person writes without a generative tool. It also does not require the planned OpenAI marking for every short output: the EU/OpenAI provenance guidance does not require watermarking for outputs shorter than 200 tokens, or for code snippets.
Who has to comply?
The obligations are divided. Providers of generative AI systems placed on the EU market must address machine-readable output marking where the rule applies. Deployers can have separate disclosure duties when they publish certain synthetic content, including relevant public-interest material. The obligations do not sit only on the company offering the system, and they do not automatically apply to every person typing a prompt.
What happens if a company ignores it?
Article 99(4) allows fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher, for certain breaches. That is a maximum subject to the legal assessment and proportionality; it does not automatically apply to every Article 50 case. The competent authority depends on the actor and the obligation involved.
OpenAI's commitment was reported on 8 October 2026. Its 2 December 2026 date is an announced implementation target, not proof that textGrain is already fully deployed. Whether a particular publication needs disclosure still depends on Article 50's provider and deployer duties, the content involved and the applicable exceptions.
