Watermarks in ChatGPT Text in the EU: What We Know and What Still Needs Confirmation

What has been announced, and what can be confirmed

Available information attributes plans to OpenAI to add invisible watermarks to text generated by ChatGPT and Codex in the European Union. However, the verified sources do not include a specific primary announcement that would allow its terms to be checked. The scope should therefore be described cautiously: there are reports of a rollout in the EU, but not enough confirmation to establish which users, models or responses would be included.

The available information also does not confirm when the rollout would begin or whether it has already been enabled for any accounts. The reports do not specify whether it would affect every ChatGPT plan, only certain products or models, or particular kinds of text. A reference to ChatGPT and Codex cannot be turned into a claim of universal coverage.

There is a specific discrepancy over the scope. One source refers to “some texts” from ChatGPT and Codex; another presents the measure as applying to all ChatGPT text in Europe. The material available does not establish whether these descriptions refer to different configurations, whether one is broader than the other, or whether either is incorrect. The responsible conclusion is that the exact coverage remains unresolved.

OpenAI has published general documentation on content provenance. That page describes an approach combining different measures and mentions watermarks based on SynthID. This is useful context about the company’s broader direction, but it does not confirm that this specific method would be used in the European text rollout described in the reports, nor does it establish the technical parameters.

The EU and the API: two rollouts that should not be conflated

The proposal distinguishes between two routes: a measure associated with ChatGPT and Codex in the EU, and an option that would be offered to API customers worldwide. Based on the verified sources available, the first can only be described as an announcement attributed to OpenAI whose scope is incomplete. The second is not sufficiently supported to claim that the option already exists, has been formally announced or is available globally.

The distinction matters because using a ChatGPT product is not the same as integrating a model through an API. In the first case, settings and features depend on the service operated by the provider. In the second, a customer can incorporate models into their own application and may need different controls. This distinction helps frame the right questions, but it does not confirm which controls OpenAI offers through either channel.

Before describing the API option as an operational feature, primary documentation would be needed to specify the supported models, how to enable it, whether the choice belongs to each customer, and what happens to text responses compared with other formats. Without those details, it is also not possible to say that watermarking is mandatory, enabled by default or removable by the customer.

What a text watermark can—and cannot—show

A watermark is intended to provide a signal that, under certain conditions, can help identify text generated by a particular system. It is not the same as a visible signature, a label at the end of a response or a record identifying the person who submitted a prompt. OpenAI’s general content-provenance documentation places watermarks within a set of measures; it is not enough to establish how a signal would be encoded in text or validated in this particular case.

Even if a detector finds a compatible signal, that alone does not establish who wrote, reviewed or published the content. Someone could have edited an AI-generated draft, combined passages from several sources or pasted the text into another document. Conversely, a detector’s failure to find a signal does not prove that the text was written entirely by a person: the text might not have been watermarked, or the signal might not have survived subsequent changes.

Editing is especially relevant to a system based on patterns within text. Rephrasing sentences, translating, summarising, correcting or combining paragraphs can change the material a detector analyses. Without technical information about the method, it is not possible to say how much editing a watermark could withstand or how much text would need to remain for it to be detected. The supported languages and minimum text length are also unknown.

Assessing reliability would require reproducible test results: a description of the protocol, the texts evaluated, languages, types of editing, thresholds, and false-positive and false-negative rates. The verified references do not provide those details. It would therefore be inappropriate to present the watermark as infallible proof of authorship or to attribute a quantified level of accuracy to it.

A product decision does not, by itself, establish a legal obligation

A company announcement and regulatory obligations are separate issues. OpenAI’s decision to add a watermark does not, by itself, establish what the AI Act requires for each category of AI-generated text. To explain a specific obligation, the relevant article, its application date, the type of system and the context of use would need to be checked against the applicable legal text.

The sources verified for this article do not include the text of the AI Act or a primary legal explanation that would resolve those questions. For that reason, this article does not attribute a specific requirement concerning this watermark to the regulation, nor does it claim that the rollout necessarily responds to a legal mandate. The available evidence supports describing a reported company initiative, not its precise legal basis.

This caution also helps avoid confusing provenance with responsibility. A watermark may provide useful information for certain review processes, but it does not, on its own, determine whether content is accurate, whether it should be published, who is responsible for it or whether all applicable obligations have been met. Those conclusions require additional facts and rules.

What remains to be learned, and how to interpret the announcement

To clarify the scope, OpenAI would need to specify publicly which models, products, languages and output types are affected; whether the measure applies to all users in the EU or only some; and the date it would begin. For the API, the information would need to clarify availability, supported models and the controls customers would receive. None of these details is resolved by the sources consulted.

A technical description is also missing if the detection process is to be assessed: what watermark is added, what conditions the detector requires, and how it performs after text is edited, translated or summarised. Without a protocol and test results, it is not possible to independently assess its robustness or the risk of misclassification.

The practical conclusion is limited but clear: reports say OpenAI plans to watermark ChatGPT and Codex text in the EU, while the exact coverage remains disputed. The verified information is not enough to confirm activation, a worldwide API option, the method used for text or the specific relationship with the AI Act. Until more specific primary documentation is available, these details should be treated as unresolved, not as established features.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.