OpenAI to Add Invisible Watermarks to EU-Generated ChatGPT Text

OpenAI has announced plans to embed invisible watermarks into text produced by ChatGPT and Codex within the European Union, responding to transparency requirements under the EU AI Act. The watermarking system will become available to eligible users across all subscription tiers in the coming weeks, though it will remain geographically limited to EU jurisdictions.

How the Watermarking Works#

Rather than inserting visible symbols, the watermark operates through subtle manipulations of the model's word selection patterns. The technique—called textGrain—subtly shapes which words the AI chooses at each step of text generation, creating a hidden signature that travels with the text even when copied or pasted. OpenAI partnered with researchers from the University of Pennsylvania and Yale to develop the approach, publishing a detailed technical report explaining how a secret key allows detectors to identify the pattern using only the final text.

Regulatory Context#

The transparency rules embedded in the EU AI Act took effect on August 2 and require AI companies to clearly mark machine-generated content in ways that other systems can reliably detect. API developers worldwide can enable watermarking for select models as of today, though OpenAI is not making it mandatory globally at launch.

Limitations and Safeguards#

Testing reveals important constraints in the watermark's durability. When researchers replaced just 10% of words with synonyms, detection accuracy fell from roughly 92% to 66%, suggesting the technique is vulnerable to editing. Short passages, mathematical solutions, and translated text present additional detection challenges. OpenAI acknowledged these shortcomings by restricting initial detector access to approved researchers and expert organizations rather than deploying it broadly.

The company also emphasized that a missing watermark should not be interpreted as proof of human authorship, since text might be too brief, heavily edited, originate from another AI system, or simply fall outside the watermark's detection range.

Broader Industry Movement#

This step follows Anthropic's announcement two months prior that it would watermark Claude's output on a global scale—a decision that faced pushback from some users who objected to having their own contributions flagged. OpenAI reportedly shelved an earlier watermarking effort over fears users would migrate to competing platforms without such detection systems. Major players including Anthropic, Google, Meta, Microsoft, and OpenAI have now committed to the EU's code of practice on identifying AI-generated material.