OpenAI to Deploy Invisible Watermark for EU ChatGPT Outputs

Compliance with the EU AI Act

OpenAI announced Monday that it will begin embedding an invisible watermark in text produced by ChatGPT and Codex for users in the European Union, aligning with the transparency provisions of the EU AI Act that took effect on August 2. The legislation requires AI providers to mark synthetic content so that other systems can recognize it.

How the watermark functions

The watermark does not appear as a visible symbol. Instead, OpenAI’s model subtly biases its word choices, creating a hidden pattern that a dedicated detector can identify. Because the signature is embedded in the words themselves, it travels with the text when it is copied or pasted. OpenAI stressed that the watermark does not reveal the user’s identity and that enabling it has not altered the model’s performance in a measurable way.

Technical report and method

Alongside the blog post, OpenAI released a technical report titled textGrain, co‑authored with researchers from the University of Pennsylvania and Yale. The document describes a technique that uses a secret key to reorder the model’s next‑word predictions, completing sentences with hundreds of tiny nudges. A detector equipped with the same key can then flag AI‑generated text using only the passage and the key.

Limitations and detection challenges

OpenAI’s internal tests show that the watermark can be weakened by editing. Replacing roughly 10 % of words with synonyms lowered detection accuracy from about 92 % to 66 %. The company also noted that short excerpts, mathematical answers, and translated material are harder to flag reliably.

“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,” OpenAI said.

What a missing or altered watermark means

OpenAI cautioned that the absence of a watermark does not prove human authorship. A passage may be too brief, heavily edited, or generated by another AI system. Conversely, a watermark only indicates that an OpenAI system generated or processed part of the text, not the extent of human judgment or creativity involved.

Industry backdrop and reactions

The move follows Anthropic’s announcement two months ago that it would apply a watermark to Claude’s output worldwide, a decision that sparked backlash from some users who felt they provided the instructions while Claude acted merely as a tool. OpenAI previously developed a watermarking system but delayed its release, partly out of fear that users might migrate to competitors that did not watermark, The Wall Street Journal reported in 2024.

Anthropic, Google, Meta, Microsoft, and OpenAI are among the signatories to the EU’s code of practice for AI‑generated content, signaling a broader industry shift toward compliance with the new regulatory framework.