OpenAI will begin adding invisible watermarks to text generated by the European Union’s ChatGPT and Codex to comply with EU AI laws, the company said in a blog post on Monday.
Transparency rules in the EU AI law, which came into force on August 2, require AI companies to mark AI-generated content in a way that other systems can identify it.
OpenAI says the watermark will roll out to eligible ChatGPT and Codex users on all plans in the coming weeks, but only within the EU. Developers using OpenAI’s APIs anywhere in the world can enable them for some models starting today. It is off by default. OpenAI said it will not make text watermarks the global default at launch.
Watermarks are not actual symbols, but work by subtly shaping the model’s word choices, leaving a pattern invisible to readers but detectable by detectors. It exists within the word itself, so it moves with the text when copying and pasting. OpenAI said that watermarks do not identify users and that turning on watermarks did not show any significant change in model performance.
At the same time as the announcement, OpenAI also released a technical report on its method called textGrain. The book, written in collaboration with researchers from the University of Pennsylvania and Yale University, describes an example of using a secret key to sort next word predictions to complete a sentence. By combining hundreds of these nudges, a detector can identify AI-generated content using only text and keys.
Can I remove the watermark by editing? OpenAI’s testing suggests that this is the case. In one test, replacing 10% of words with synonyms reduced the detection rate from about 92% to 66%. The company also said short passages, math answers, and translated text are difficult to detect.
“These limitations contribute to our decision to provide access to the initial detector only to approved researchers and professional organizations who can help assess its reliability and responsible use,” the company said.
OpenAI also warned that missing watermarks “cannot prove human authorship.” The text may be too short or heavily edited. Or it could be from another company’s AI.
“[Watermarks]can indicate that an OpenAI system generated or processed a portion of the text, but they cannot indicate the extent to which human judgment, editing, or creativity was involved,” the company said.
The announcement comes two months after Anthropic announced it would watermark text generated by Claude, an initiative that is being applied around the world. The decision prompted a backlash from some Claude users, who argued that Claude was only a “tool” and that they provided “instruction, context, and decisions.”
The Wall Street Journal reported in 2024 that OpenAI had previously built a text watermark, but held off on releasing it due to concerns that users would switch to rivals that don’t have watermarks.
Anthropic, Google, Meta, Microsoft, and OpenAI are companies committed to following the EU Code of Conduct for AI-generated content.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
