Hajra Naz
OpenAI will start adding an invisible watermark to text. This applies to content generated by ChatGPT and Codex. The move specifically targets the European Union. OpenAI announced this in a blog post on Monday. The goal is compliance with the EU AI Act.
The EU AI Act’s transparency rules took effect on August 2. These rules require AI companies to mark AI-generated content clearly. The marking needs to be identifiable by other systems.
OpenAI said this watermark will roll out over the coming weeks. It applies to eligible ChatGPT and Codex users across all plans. This rollout is limited to the EU, though. Developers using OpenAI’s API face different rules. They can turn on watermarking for select models starting today, anywhere in the world. It remains off by default, though. OpenAI confirmed it’s not making text watermarking a global default at launch.
Read More: ChatGPT Will Show Visual Ads During Image Generation, Starting With a US Test
This watermark isn’t an actual visible symbol. Instead, it works by subtly shaping the model’s word choices. This leaves a pattern readers can’t see. A detector can pick it up, though. Because the watermark lives within the words themselves, it travels with the text. This happens even when the text gets copied and pasted elsewhere. OpenAI said the watermark doesn’t identify specific users. The company also saw no meaningful change in model performance with watermarking switched on.
OpenAI also published a technical report detailing its method. This approach is called textGrain. The report was co-written with researchers from the University of Pennsylvania and Yale. It walks through a specific example. A secret key sorts next-word predictions to complete a sentence. Combine hundreds of these small nudges, and detection becomes possible. A detector can spot AI-generated content using just the text and the key.
Can editing remove this watermark? OpenAI’s own tests suggest yes, in some cases. In one test, replacing 10% of words with synonyms had a significant effect. Detection accuracy dropped from about 92% to 66%. The company also noted other challenges. Short passages, math answers, and translated text prove harder to detect accurately.
Read More: OpenAI Safety Employee Resigns, Says Company Culture Is Broken
“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 said.
OpenAI also offered an important caution. A missing watermark “does not prove human authorship.” Several explanations could account for this. The text might be too short. It could be heavily edited. It might also come from a different company’s AI system entirely.
“[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,” the company said.
This announcement comes two months after a similar move from Anthropic. That company said it would watermark text generated by Claude. Notably, Anthropic applied that watermarking worldwide. That decision drew real backlash from some Claude users. Many argued they had supplied “the instructions, context, decisions.” They felt Claude functioned merely as “the tool” in that process.
Read More: OpenAI’s GPT-6.1 Sol Cuts Frontier AI Pricing to a Fifth of Astra’s
OpenAI had actually built a text watermark before this announcement. The company held off on releasing it previously. Part of the concern involved users switching to rival platforms. Specifically, platforms that didn’t watermark their content at all. The Wall Street Journal reported this detail back in 2024.
Several major companies have committed to following the EU’s code of practice on AI-generated content. That includes Anthropic, Google, Meta, Microsoft, and OpenAI.
