OpenAI Watermarking Analysis: EU ChatGPT, API Opt In

A compliance deadline rarely arrives as a grand announcement. More often, it appears as a product setting that is on for one group, off for another, and surrounded by documentation lawyers will pretend is obvious. According to TechCrunch, OpenAI will add an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act. The important part for builders is the split: eligible ChatGPT and Codex users in the EU get the default, while API developers globally get an opt in setting for select models.

The rule lands as a product default, according to TechCrunch

TechCrunch reports that the EU AI Act transparency rules took effect on August 2 and require AI companies to mark AI generated content in a way other systems can identify. OpenAI said the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, but only in the EU. Developers using OpenAI's API anywhere in the world can turn it on for select models starting today, and it is off by default. OpenAI also said it is not making text watermarking a global default at launch, which is the sentence your implementation team should underline, not politely nod past.

@title OpenAI watermark rollout
@source OpenAI will start watermarking ChatGPT's text in the EU
@source Our approach to EU text provenance rules

  August 2 ··· EU rules apply
  October 5 ·· OpenAI post
  October 5 ·· API option starts
  Coming weeks · EU rollout

@caption The compliance change arrives first as geography and default settings.

That is the practical shape of the obligation. For ordinary ChatGPT and Codex users in the EU, the marking happens inside the product once the rollout reaches them. For API customers, there is no magic compliance fairy turning the feature on worldwide. If an app serves EU users and relies on OpenAI's API, the first implementation question is not whether watermarking exists, but whether the developer has enabled it for the relevant model and use case.

The mark is in the wording, not on the page, according to OpenAI

OpenAI describes the effort on its own site as an approach to EU text provenance rules, with the aim of promoting transparency within current technical limits. TechCrunch reports that the watermark is not a visible symbol. Instead, it works by subtly shaping the model's word choices, leaving a pattern that readers cannot see but a detector can identify. Because the pattern lives in the words themselves, TechCrunch says it travels with the text when copied and pasted.

OpenAI also published a technical report for the method, called textGrain, according to TechCrunch. The report was co written with researchers from the University of Pennsylvania and Yale, and describes using a secret key to sort next word predictions while completing a sentence. Over many such nudges, the detector can identify AI generated content using only the text and the key. This is not a user tracking label: TechCrunch reports that OpenAI said the watermark does not identify the user, and that it saw no meaningful change in model performance with the feature switched on.

What the law requires is narrower than the discourse, according to 9to5Mac

9to5Mac frames the move as OpenAI explaining how it will comply with EU AI Act requirements for providers of generative AI systems to mark AI generated text in a way that makes it identifiable. That does not mean every sentence generated everywhere by every OpenAI product is now visibly branded. It means a provider is implementing a machine identifiable marking mechanism for the regulated surface it has chosen to address first. LinkedIn will turn that into twelve compliance morals by lunch; the operative facts are smaller and more useful.

The distinction matters because the same model provider is now offering different behavior by product surface and geography. ChatGPT and Codex in the EU are treated differently from API use elsewhere. Even within API use, the setting is available for select models and off by default, according to TechCrunch. If you are drafting customer terms, procurement questionnaires, or school policies, do not write the lazy sentence that OpenAI text is watermarked. Write which product, which geography, which model, and whether the setting is enabled.

The builder checklist is mostly about defaults, according to TechCrunch and OpenAI

For product teams, the immediate compliance exercise is inventory, not philosophy. Map where your application calls ChatGPT, Codex, or the OpenAI API, then separate EU end user surfaces from non EU ones. If you use the API, TechCrunch's reporting means you should not assume watermarking is active unless your configuration says so. If you resell or embed generated text, your contracts should say who is responsible for turning on available provenance features and how that setting is verified.

For educators and workplace policy teams, the caveat is equally plain. A watermark detector can support provenance checks, but TechCrunch's description does not make it a universal authorship oracle. OpenAI says the watermark does not identify the user, and the mechanism depends on detecting a pattern in text. Treat it as one signal about whether text likely came from a marked generation system, not as a substitute for assessment design, disclosure rules, or basic evidence.

The next thing to watch is whether other providers follow the same segmented pattern or choose wider defaults. The EU AI Act is starting to show up not as one global switch, but as product surface segmentation: EU consumer tools one way, API customers another, and the rest of the world waiting on a business decision. Builders should document those seams now. Regulators tend to enjoy seams almost as much as I enjoy enforcement notices.