OpenAI announced on Monday via a blog post that it will introduce an invisible watermark to text produced by ChatGPT and Codex within the European Union, as a means of complying with the EU AI Act. The transparency requirements of the EU AI Act, which came into force on August 2, mandate that AI companies label AI-generated content in a manner detectable by other systems.

According to OpenAI, the watermark will be deployed over the next several weeks to eligible ChatGPT and Codex users across all subscription plans, though exclusively within the EU. Developers who use OpenAI's API worldwide will have the option to enable it for certain models beginning today; however, it remains disabled by default. OpenAI stated that it is not making text watermarking a default setting globally at this time.

Rather than being a visible symbol, the watermark functions by subtly influencing the model's selection of words, creating a pattern that is imperceptible to readers but detectable by specialized tools. Since it is embedded within the words themselves, the watermark persists when text is copied and pasted. OpenAI emphasized that the watermark does not reveal user identity and that its models showed no significant performance degradation when the feature was active.

Alongside the announcement, OpenAI released a technical report detailing its method, which it calls textGrain. The report, co-authored with researchers from the University of Pennsylvania and Yale, provides an example of how a secret key can be used to rank next-word predictions to complete a sentence. By accumulating hundreds of these subtle adjustments, a detector can identify AI-generated content using only the text and the key.

When it comes to whether editing can strip the watermark, OpenAI's testing indicates that it can. In one experiment, substituting 10% of words with synonyms reduced detection accuracy from approximately 92% to 66%. The company also noted that short passages, mathematical answers, and translated text are more difficult to detect.

"These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses," the company stated. OpenAI further warned that the absence of a watermark "does not prove human authorship." The text might be too brief, too extensively modified, or generated by a different company's AI system. "[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it," OpenAI explained.

This announcement follows two months after Anthropic revealed plans to watermark text produced by Claude, a policy it is implementing on a global scale. That move sparked criticism from some Claude users, who contended that they had provided "the instructions, context, decisions" while Claude served merely as "the tool."

OpenAI had previously developed a text watermark but chose not to release it, partly due to worries that users might migrate to competitors that did not employ watermarking, according to a 2024 report. Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have pledged to adhere to the EU's code of practice concerning AI-generated content.