OpenAI Adds an Invisible Text Watermark to ChatGPT in the EU

OpenAI said on Monday, October 5 that it will add an invisible watermark called textGrain to eligible ChatGPT and Codex text output in the European Union over the coming weeks. OpenAI said the move responds to EU AI Act requirements that generated text be identifiable in a machine readable way. The company also said text watermarking and detection are still early technologies with significant limits, so results need careful interpretation by teachers, editors and readers in the EU.

OpenAI said textGrain works by adding an invisible statistical signal to the word choices made by a model. A detector then looks for that signal and assesses whether a passage contains an OpenAI watermark. The signal is not a visible label in the writing. OpenAI said approved researchers and expert organizations can apply for detector access at launch, while the detector will not be made publicly available at first because missed watermarks and false positives remain possible.

OpenAI said a detector result can be wrong in either direction. It can report a watermark where none is present, which OpenAI calls a false positive, or miss a watermark that is present, which it calls a false negative. In evaluations described by OpenAI, detection was stronger for longer passages and weaker for shorter or more constrained writing. OpenAI also said performance under ideal conditions does not guarantee reliable detection in everyday use, which limits what any single result can establish.

For a teacher, editor or reader, OpenAI said a detected watermark does not measure how much a person contributed to a piece of writing. It can indicate that an OpenAI system generated or processed part of a passage, but it cannot show how much human judgment, editing or creativity was involved. OpenAI also said a watermark does not establish ownership or responsibility, decide whether use was lawful, determine whether disclosure was required, or identify who is responsible for the text.

OpenAI also said a watermark does not identify the user behind a passage. It does not associate the text with a person, organization, account, prompt or conversation. The company said a watermark does not verify whether the writing is accurate, true, misleading, harmful or presented in the right context. That means a positive result is a signal about possible OpenAI processing, not a verdict about authorship, quality, truth, ownership or responsibility, according to the company.

OpenAI said the absence of a detected watermark does not prove that a human wrote the text. Detection may not work reliably when text is too short, edited or translated, and the company said text may also come from an unsupported model, may predate watermarking, or may have been generated by tools from another company. In the same way, a positive result can still be a false positive. The Verge, reporting the launch on October 5, noted that OpenAI itself says textGrain does not guarantee reliable detection.

OpenAI said the EU rollout will cover eligible ChatGPT and Codex users across all plans, while API customers globally can opt in to watermarking for select models and the API setting will remain off by default. The company said it plans to revisit the approach as technology, standards and evidence develop. For students and working writers, the practical point is that honest disclosure and keeping drafts and notes remain the dependable way to show how a piece of writing was made.