OpenAI is rolling out invisible, machine-readable watermarks on text generated by ChatGPT and Codex, but for now the feature is limited to users in the European Union. The new system, called textGrain, embeds a detectable pattern into AI-written text without changing how it reads, and OpenAI says it was built largely to satisfy requirements under the EU AI Act.
What textGrain Actually Does
Unlike visible watermarks on images, textGrain works invisibly inside the words themselves. OpenAI claims the technique “matched or exceeded” rival approaches, including Google DeepMind’s SynthID for text, which is also the basis for the watermarking Anthropic announced in August. The company published benchmark scores showing that watermarked and unwatermarked outputs perform similarly, meaning the watermark shouldn’t noticeably degrade the quality of responses from ChatGPT or Codex.
Crucially, OpenAI is upfront about the limits of the technology. The company says textGrain “does not guarantee reliable detection,” and it cannot confirm who owns a piece of text, verify its accuracy, measure how much a human edited it, or prove that a human wrote any of it at all. In other words, this is a transparency tool, not a lie detector or plagiarism shield.
Why the EU Gets It First
OpenAI is explicit that this rollout is regional, not global. Over the coming weeks, eligible ChatGPT and Codex users across all plans in the EU will start seeing watermarked text, but OpenAI says it is “not making text watermarking a global default at launch.” The company frames this as a way to learn from real-world feedback before deciding whether to expand further.
That caution lines up with the timing: the EU AI Act includes transparency obligations for AI-generated content, and Anthropic made its own move for the same regulatory reason in August. Both companies are essentially testing watermarking in the one market where it’s currently required, rather than committing to it everywhere.
API Access and Detector Rules
Outside the EU, OpenAI is taking a lighter touch. Starting now, API customers anywhere in the world can opt in to watermarked outputs for select models, letting businesses decide whether watermarking fits their own compliance needs or the experience they want to offer users. OpenAI also says it’s working with cloud partners to extend watermarking to its models accessed through third-party cloud services in the coming weeks.
Detection, meanwhile, remains tightly controlled. Approved researchers and expert organizations can apply for access to a detector tool starting today, but OpenAI is granting access case by case, in line with the EU’s Code of Practice. The detector only reports whether a watermark is present, without exposing the user’s identity or their conversation. OpenAI says it’s avoiding a public release of the tool because of the risk of missed watermarks and false positives, both of which could wrongly accuse someone of using (or not using) AI.
The Bottom Line
This is a cautious, compliance-driven first step rather than a sweeping anti-cheating or anti-misinformation measure. For everyday ChatGPT users outside the EU, nothing changes yet, and even EU users won’t notice any visible difference in their output. The real story is that OpenAI is testing how invisible watermarking performs at scale before deciding whether to push it worldwide, while repeatedly reminding everyone that a watermark alone can’t prove who wrote what.
Source: The Verge
