On October 5, 2026, OpenAI announced plans to add invisible watermarks to eligible text generated by ChatGPT and Codex in the European Union. API customers worldwide can opt in for selected models, while the watermark detection tool will initially be available only to approved researchers and specialist organizations.
The notable point is not just that AI-generated text will gain an identifying signal. OpenAI also emphasizes that detection results are not enough to determine who wrote the content, how much humans contributed or whether the content is accurate.
EU rollout, with the API still off by default
In its announcement on its approach to EU text provenance requirements, OpenAI frames the rollout in the context of the EU AI Act's requirement for AI-generated text to be identifiable through machine-readable means.
Under the announced plan, watermarks will be added over the following weeks to eligible text output from ChatGPT and Codex in the EU, across all plans. This is not a global default-on launch.
For the API, customers worldwide can choose to enable watermarking for selected models from the day of the announcement. The feature remains off by default, leaving application developers to decide how to integrate it into their products and transparency requirements.
This scope needs to be understood correctly: the announcement does not say that all text from every OpenAI model will carry a watermark. Therefore, failing to find a signal does not mean the text was definitely written by a human.
textGrain identifies statistical signals, not accounts
The technology OpenAI calls textGrain adds an invisible statistical signal to the way a model selects words. The detector looks for that signal to assess whether a passage contains an OpenAI watermark.
This is neither a visible label on the page nor a user identifier. According to OpenAI, the tool does not link text to the individual, organization, account, prompt or conversation that produced it.
This distinction matters for businesses using AI to draft documents or marketing content. A detection result may indicate that an OpenAI system generated or processed part of a passage, but it cannot measure the extent of the content creator's editing, creativity or judgment.
A watermark also does not establish ownership, the legality of its use or who is responsible. It provides a signal of origin, not a replacement for editorial records and approval procedures.
Short text and editing reduce detectability
OpenAI published evaluations showing that text length and content type significantly affect detection performance. At a target false-positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages, for content such as psychology.
A token is a unit a model processes, not a fixed equivalent of one word. These results should therefore not be converted into a minimum Vietnamese word count that guarantees detection.
For content such as mathematics, detection rates are significantly lower because the model has less flexibility in its word choices. The figures above are evaluation results published by OpenAI, not guarantees for every language or every real-world document.
In another evaluation using 400-token English passages, replacing 10% of the words with synonyms reduced the detection rate from about 92% to 66%. Replacing 25% brought the rate down to 17%.
This shows that even changes to wording can weaken the signal. OpenAI also cites translation as a possible reason why a watermark may not be reliably detected. Conversely, the detector can still report a watermark in text that does not contain one—a false positive.
Why isn't the detector open to the public yet?
OpenAI says initial access is considered on a case-by-case basis for researchers and specialist organizations, to evaluate reliability and responsible use. The company has not opened the text detector to the public at launch because of the risk of missed watermarks and incorrect detections.
This restriction applies only to the text verification tool in the announcement. The image and audio verification tools mentioned by OpenAI remain publicly accessible.
For ordinary users, the new watermark detector therefore cannot yet be treated as a public utility for checking any article. Even with access, its results should not be used on their own as decisive evidence of authorship or cheating.
Provenance transparency still needs its own process
For content teams and businesses, this change adds a layer of technical signals, but it does not replace recording sources, checking facts and identifying the final approver. This topic is directly related to the website's AI content transparency policy: disclosing how AI is used is a separate issue from being able to detect signs of machine-generated text.
A watermark is not a measure of quality, either. Text with an OpenAI signal can still be incorrect or lack context; text without one can still be AI-generated. The process of writing prompts with clear requirements and checking the results therefore remains necessary, whether or not a watermark is present.
For application developers, the option to enable watermarking through the API needs to be considered alongside output requirements and operational workflows, rather than treated as a switch that solves the entire transparency problem. This also extends the issue of testing and oversight when deploying AI.
OpenAI expects to review the rollout's scope and access as the technology, standards and real-world evidence change. The company also announced plans to make the technology open source, but the announcement did not specify a timeline. The evaluations cited currently provide insufficient grounds to conclude how reliably Vietnamese text can be detected.