News about “invisible watermarks” in AI text is especially easy to misread when regulatory requirements are conflated with a vendor’s product plans. The EU has established that some AI-generated content must be identifiable, and it requires clear, visible disclosure in certain situations. But the body of OpenAI’s official page titled “EU text provenance” was still inaccessible at press time, so there is currently no basis for confirming whether OpenAI has decided which ChatGPT or Codex text would receive what kind of mark.
That distinction matters to teams using AI to write customer-service scripts, code documentation, and marketing copy. The compliance direction can be understood, but OpenAI’s implementation cannot be filled in from a headline or speculation. Mixing the two can easily turn a possible plan into a feature that has already launched.
The EU already requires AI-generated content to be identifiable
The European Commission’s explanation of the Artificial Intelligence Act (AI Act) is clear: providers of generative AI must ensure that AI-generated content is identifiable. Deepfakes, as well as certain AI-generated texts published to inform the public about matters of public interest, must also carry clear and visible disclosure.
These transparency rules will apply from August 2, 2026. They focus on whether content can be identified and whether certain content is explicitly disclosed to the public. The overview does not specify that every company must use one particular technology.
In other words, the law sets the required outcome first; it does not mean that statistical watermarking is the mandated way to achieve it. A company may use machine-readable marks, detection mechanisms, or another compliance approach, but it still has to show that its measures are effective enough and fit the applicable context.
The European Commission has published a Code of Practice on Transparency of AI-Generated Content. It covers two types of work. For providers, it addresses machine-readable marking of audio, images, video, and text outputs, so that content can be identified as generated or manipulated by AI. For deployers, it addresses public disclosure for deepfakes and certain AI-generated texts.
One distinction should not be omitted: the code is a voluntary compliance tool, while the transparency requirements in Article 50 are legal obligations. The Commission says providers and deployers that sign and follow the code can use its measures to demonstrate compliance. Companies that choose other methods must show that their measures are equally appropriate. The code does not replace the AI Act or the Commission’s guidance on the transparency obligations.
The scope of OpenAI’s plan remains unconfirmed
The title of OpenAI’s page points to “EU text provenance,” but its body cannot currently be read from the public page. As a result, the following key details cannot be treated as confirmed product facts: whether the plan applies to ChatGPT, Codex, or other products; whether its scope is defined by region, model, language, or text type; whether it is in research, testing, planning, or deployment; and how outside organizations would detect it or whether researchers could access related tools.
In particular, “some text” must not be expanded into “all ChatGPT responses,” nor should it be rewritten as “all Codex code will be marked.” That single qualifier could determine whether this is a limited compliance arrangement or a product update that changes companies’ everyday workflows.
It is also impossible at present to say which outputs would be marked, whether a mark would survive copying and editing, or whether the system would provide traceability. Whether short text, translated text, human-edited text, and human-AI mixed text fall within the same detection scope requires publicly verifiable explanations or test results from OpenAI. It cannot be inferred from EU law.
That is why the term “invisible watermark” should be used carefully. It is at least different from writing “AI-generated” directly on an article page, and it is not necessarily the same as file metadata. But without the body of OpenAI’s page, there is no basis for asserting which specific mechanism it uses, much less for estimating its accuracy, resistance to rewriting, or false-positive rate.
Product teams should not put the headline into a requirements document yet
For now, companies can incorporate the EU’s transparency obligations into their content-governance plans: identify which AI-generated content needs to be identifiable, determine which public-facing text might trigger visible disclosure, and retain a verification step for vendors’ compliance statements.
But until OpenAI publishes an accessible official explanation, companies should not use this as the basis for changing user agreements, promising customers that “all outputs have watermarks,” or documenting a detection capability as already available. Platforms and media outlets should also distinguish between “possibly generated by AI” and “whether the content is true.” The former is a source-identification question; the latter still requires fact-checking.
The most defensible conclusion is simple: the EU’s transparency requirements for AI-generated content are clear, but OpenAI’s scope and method for applying them to ChatGPT, Codex, or other text outputs still cannot be confirmed from this official page, whose body remains inaccessible. Before integrating or procuring such a system, ask the vendor for three things: its scope, information about detection and false positives, and whether the mark remains effective after editing or rewriting.
Sources
- European Commission: AI Act
- European Commission: Code of Practice on Transparency of AI-Generated Content
- European Commission: Article 50 of the AI Act
- OpenAI: EU text provenance (body inaccessible at press time; product details were not confirmed from it)
