OpenAI is beginning to roll out text watermarking in ChatGPT and Codex, adding an invisible, machine-readable mark to text output. The move is starting with users in the European Union, according to the reporting, and appears tied to broader compliance efforts around the EU AI Act. The development puts a new layer of provenance-related identification into two of OpenAI’s text-generation products while keeping the change out of view for users reading the output itself.
The reported rollout is notable because it concerns text rather than images or audio. In this case, the watermark is meant to live inside the generated words in a way that machine systems can later detect. That makes textGrain different from a visible label or disclaimer, since the goal is not to warn readers directly but to create a machine-readable signal that can support identification of generated text. The feature applies to ChatGPT and Codex, and the rollout is limited to EU users at first.
What OpenAI is adding
The system is described as a text watermark called textGrain. Unlike a visible label, this watermark is meant to be machine-readable and embedded in generated text. The change applies to ChatGPT and Codex, and the rollout is limited to EU users at first. In other words, the same generation tools remain available, but the output is now being marked in a way that is designed for automated recognition rather than for human attention.
According to the source material, OpenAI says textGrain “matched or exceeded” other approaches such as Google DeepMind’s SynthID for text. The same report notes that Anthropic’s watermarking work, announced in August, is also based on SynthID for text. Those comparisons place OpenAI’s approach alongside other efforts that are trying to make generated text more traceable in ways that can be checked by systems, not just by people.
Why the change matters
The reporting links the addition to requirements connected with the EU’s AI Act. OpenAI is not the only company moving in this direction: Anthropic also made a similar move to meet those requirements. That places text watermarking inside a wider industry pattern of preparing generated text for identification or provenance checks. The rollout therefore reads as both a product update and a compliance response, rather than as a standalone feature with no policy context.
OpenAI also shared benchmark scores that showed similar performance between watermarked and unwatermarked text. That detail suggests the company is aiming to add the watermark without noticeably changing output quality, at least in the tests it included. The reported benchmark result is important because it frames the watermark as something meant to be unobtrusive in use, even while it changes how the text can be handled by detection systems later on.
Limits of the system
The source material also makes clear that textGrain is not a perfect safeguard. OpenAI notes that the system “does not guarantee” that every piece of generated text can be identified, which means the watermark is intended as a detection aid rather than an absolute proof of origin. That distinction matters because the presence of a watermark can strengthen provenance checks, but it does not settle every question about where text came from or whether it can be reliably detected in all cases.
That limitation is important because invisible text watermarking depends on consistent detection tools and implementation. In practice, a watermark can support authenticity checks, but it does not eliminate uncertainty on its own. The reporting therefore presents textGrain as a useful mechanism with clear bounds, not as a complete answer to the problem of identifying machine-generated writing.
Related signals from OpenAI
Separate reporting from the same day says OpenAI is testing more ads in ChatGPT. While that development is distinct from text watermarking, it shows the company is continuing to adjust the platform in multiple ways at once. The ad testing report does not change the details of the watermarking rollout, but it adds to the picture of a product that is still actively evolving. Taken together, the two reports show OpenAI making changes both to how content is surfaced and to how it may be identified after generation.
Conclusion
OpenAI’s textGrain rollout represents another step toward machine-readable identification of AI-generated text. For now, the feature is limited to ChatGPT and Codex users in the European Union, and OpenAI says it performs similarly to unwatermarked text in the benchmarks it shared. At the same time, the company acknowledges the watermark is not a guarantee, which means the system is best understood as one layer of detection rather than a complete solution.
The broader significance of the move is that it fits into a wider trend around text provenance and compliance. With OpenAI and Anthropic both moving in ways connected to the EU AI Act, the reporting suggests that invisible watermarking is becoming part of the standard toolkit for companies that want to identify generated text without changing the experience of reading it. In OpenAI’s case, textGrain is being introduced as a machine-readable layer beneath the output, not as a visible badge on top of it.