OpenAI to launch invisible text watermarking in EU

OpenAI has announced that it will begin adding an invisible watermark to text generated by ChatGPT and Codex within the European Union. The company stated on Monday that this measure is required to comply with the transparency rules of the EU AI Act. The watermark is designed to allow other systems to identify AI-generated content, a requirement that took effect on August 2. This initiative marks a significant step in regulatory compliance for artificial intelligence providers operating in the region.

The rollout is scheduled to occur over the coming weeks and will apply to eligible users of ChatGPT and Codex on all subscription plans, but strictly within the EU. Developers who utilise OpenAI’s API anywhere in the world can enable the feature for select models starting immediately, although it remains off by default. OpenAI clarified that it is not making text watermarking a global default at launch. The company emphasised that the watermark does not identify the user and reported that it saw no meaningful change in its models’ performance with the feature switched on.

The technology, known as textGrain, works by subtly shaping the model’s word choices to leave a pattern that is invisible to readers but detectable by specific software. Because the watermark is embedded in the words themselves, it travels with the text when it is copied and pasted. OpenAI published a technical report detailing this method, which was co-written with researchers from the University of Pennsylvania and Yale. The report explains how a secret key is used to sort next-word predictions, allowing a detector to spot AI-generated content using only the text and the key.

However, the company acknowledged limitations in the technology’s durability. Tests suggest that the watermark can be removed by editing the text. In one specific test, replacing 10% of words with synonyms dropped detection rates from about 92% to 66%. OpenAI also noted that short passages, mathematical answers, and translated text are harder to detect. Consequently, the company is providing initial detector access only to approved researchers and expert organisations to help evaluate reliability and responsible uses.

OpenAI cautioned that a missing watermark does not prove human authorship, as the text could be too short, heavily edited, or generated by another company’s AI. The company stated that watermarks can indicate that an OpenAI system generated or processed part of a passage, but they do not reveal how much human judgment, editing, or creativity went into the final output. This announcement follows a similar move by Anthropic two months ago, which decided to watermark text generated by Claude worldwide, a decision that drew backlash from some users who argued they supplied the instructions and context while the AI was merely the tool.

OpenAI had previously built a text watermark but held off on releasing it, partly due to concerns that users would switch to rivals that did not watermark, according to a report by The Wall Street Journal in 2024. Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to following the EU’s code of practice on AI-generated content. The new watermarking system represents a broader industry effort to address transparency and accountability in the deployment of generative AI technologies.