OpenAI has announced plans to embed invisible watermarks into text produced by ChatGPT and its Codex coding tool for users located within the European Union. The move is a direct response to compliance requirements under the EU AI Act, which the company confirmed in a blog post published Monday.

Transparency provisions within the EU AI Act came into force on August 2, placing an obligation on artificial intelligence companies to mark AI-generated content in a manner that other systems are capable of identifying. OpenAI's watermarking rollout represents one of the more concrete steps taken by a major AI provider to meet those requirements.

How the Rollout Works

According to OpenAI, the watermark will be introduced gradually over the coming weeks for eligible ChatGPT and Codex users across all subscription tiers, though the initial deployment is limited to the EU. Developers who access OpenAI's services through its API, regardless of where they are located in the world, can opt into the feature for certain models starting immediately. However, the feature is switched off by default, and OpenAI has stated it has no plans to make text watermarking a global default at this stage of the launch.

What the Watermark Actually Does

The watermark is not a visible symbol or tag appended to the text. Instead, it functions by subtly influencing the model's word selection during generation, embedding a pattern within the language itself that is imperceptible to human readers but detectable by a purpose-built system. Because the watermark is woven into the words rather than attached as metadata, it remains present when text is copied and pasted into other documents or platforms.

OpenAI has clarified that the watermark does not contain any information that could identify the individual user who generated the content. The company also reported that enabling the watermark produced no meaningful degradation in the quality or performance of its models.

The Technical Method Behind It

Alongside the announcement, OpenAI released a technical report detailing the underlying methodology, which the company has named textGrain. The report was co-authored with researchers from the University of Pennsylvania and Yale University. It explains how a secret key is used to sort and rank next-word predictions during text generation, introducing slight but consistent biases in word choice. When hundreds of these small nudges accumulate across a passage, a detection system can identify the content as AI-generated using only the text itself and the corresponding key.

Known Limitations

OpenAI's own testing indicates that the watermark can be disrupted through editing. In one assessment, replacing just 10% of words in a passage with synonyms caused detection accuracy to fall from approximately 92% to 66%. The company also acknowledged that certain types of content are inherently more difficult to watermark reliably, including short passages, mathematical answers, and text that has been translated from one language to another.

Given these constraints, OpenAI has chosen to restrict access to its detection tool during the initial phase. Rather than making it broadly available, the company will provide access only to approved researchers and expert organizations that can assess the system's reliability and help define appropriate use cases.

"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 was also careful to note that the absence of a watermark should not be interpreted as evidence that a piece of text was written by a human. A passage might lack a detectable watermark because it is too short, because it has been heavily edited after generation, or because it was produced by a different company's AI system entirely.

The company further noted that even when a watermark is present, it does not convey the degree to which human judgment, editing, or creative input shaped the final output. A watermark can indicate that an OpenAI system was involved in generating or processing part of a passage, but it cannot quantify the human contribution.

Industry Context

The announcement follows a similar move by Anthropic, which disclosed roughly two months ago that it would apply watermarks to text generated by its Claude AI assistant. Unlike OpenAI's EU-specific rollout, Anthropic's implementation applies globally. That decision prompted criticism from a portion of Claude's user base, with some users arguing that they had provided the instructions, context, and decision-making while the AI itself was merely functioning as a tool.

OpenAI had developed a text watermarking capability prior to this announcement but chose not to release it. Reporting from The Wall Street Journal in 2024 indicated that one of the key reasons for the delay was concern that deploying watermarks could push users toward competing AI products that had not implemented similar measures.

OpenAI is among a group of major technology companies - including Anthropic, Google, Meta, and Microsoft - that have formally committed to adhering to the EU's code of practice on AI-generated content, which outlines voluntary standards for transparency and responsible deployment of AI systems across the bloc.