OpenAI will begin embedding an invisible watermark into text generated by ChatGPT and Codex for users in the European Union, a move designed to align with the bloc’s sweeping AI transparency rules.

The company said the rollout will take place over the coming weeks for eligible ChatGPT and Codex users across all plans, but only within the E.U. Developers using OpenAI’s API anywhere in the world will also be able to enable the feature for select models starting today, though it will remain off by default. OpenAI is not making text watermarking a global default at launch.

Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram

How The Watermark Works

The watermark is not a visible mark or symbol. Instead, it works by subtly influencing the model’s word choices in a way that leaves a detectable pattern hidden in the text itself. Readers will not notice it, but a compatible detector can identify it. Because the signal is embedded in the language, it remains attached even when the text is copied and pasted.

OpenAI said the system does not identify individual users and does not appear to materially affect model performance when enabled.

Built For The E.U. AI Act

The move is aimed squarely at compliance with the E.U. AI Act’s transparency obligations, which took effect on August 2 and require AI companies to mark AI-generated content in a way that other systems can detect.

That regulatory pressure is reshaping product decisions across the industry. For leading AI developers, transparency is no longer a theoretical policy question; it is now a design constraint.

Inside TextGrain, OpenAI’s Watermarking Method

Alongside the announcement, OpenAI published a technical report on the method, called textGrain. The paper, co-authored with researchers from the University of Pennsylvania and Yale, describes a technique that uses a secret key to influence next-word predictions. Over many such adjustments, the pattern becomes detectable by a system that has access to both the text and the key.

The approach is designed to flag likely AI-generated passages without adding any visible markers for readers.

Limits Remain A Core Concern

OpenAI also acknowledged that the system is not foolproof. In internal testing, replacing 10% of words with synonyms reduced detection accuracy from about 92% to 66%. The company said short passages, math-heavy answers, and translated text are also more difficult to detect.

Those limitations are why OpenAI said it will initially provide detector access only to approved researchers and expert organizations, which it believes are best positioned to evaluate reliability and responsible use.

The company further cautioned that the absence of a watermark does not prove human authorship. A passage may be too short, too heavily edited, or generated by another company’s model. As OpenAI put it, the watermark can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity was involved.

A Broader Industry Shift Toward Provenance

The announcement follows Anthropic’s decision to watermark text generated by Claude, a policy it is applying globally. That move drew criticism from some users, who argued they were providing the direction, context, and judgment while the model simply served as a tool.

OpenAI had previously built a text watermark but did not release it, reportedly in part because of concerns that users would migrate to rival systems that did not impose the same disclosure standards.

Now, with the E.U. drawing a firmer line on AI transparency, provenance is becoming a competitive and compliance issue at once. Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to the E.U.’s code of practice on AI-generated content, signaling that watermarking and content attribution are likely to become a defining feature of the next phase of the AI market.