OpenAI announced it will embed an invisible watermark into text produced by ChatGPT and Codex for users in the European Union, a step taken to satisfy the transparency obligations of the EU AI Act that became enforceable on August 2. The new measure aims to make AI-generated content identifiable by downstream systems.
The rollout will begin in the coming weeks for all eligible ChatGPT and Codex accounts on every subscription tier within the EU. Developers who access OpenAI’s API from any location can activate the watermark for selected models starting today, though it remains disabled by default and will not be applied globally at launch.
The watermark does not appear as a visible symbol; instead it subtly steers the model’s word selection, embedding a statistical pattern that human readers cannot perceive but that a detector can extract. Because the pattern is encoded in the words themselves, it persists when the text is copied or pasted, and it does not reveal the identity of the user or degrade model performance.
OpenAI accompanied the announcement with a technical paper titled textGrain, co-authored with scholars from the University of Pennsylvania and Yale. The report details a method that uses a secret key to rank next-word probabilities, applying hundreds of tiny nudges that together create a detectable signature. Detection can be performed using only the generated text and the shared key, without external metadata.
Tests show the watermark can be weakened by editing; substituting roughly ten percent of the words with synonyms reduced detection rates from about ninety-two percent to sixty-six percent. The company also noted that brief excerpts, mathematical answers, and translated passages present greater challenges for the detector. Consequently, OpenAI will initially grant detector access only to vetted researchers and expert organizations to assess reliability and responsible use.
OpenAI warned that the absence of a watermark does not prove human authorship, as short or heavily edited text, or content generated by another AI system, may also lack the marker. While the watermark signals that an OpenAI model contributed to a passage, it does not quantify the amount of human judgment, editing, or creativity involved. The move follows Anthropic’s recent decision to watermark Claude globally, a step that sparked user backlash over perceived tool ownership.