OpenAI to Add Invisible Watermarks to AI-Generated Text in Europe

OpenAI has announced that it will begin adding invisible watermarks to eligible text generated by ChatGPT and Codex in the European Union. The move is part of the company’s response to the EU AI Act, which requires providers of generative AI systems to make AI-generated content identifiable in a machine-readable way. OpenAI says the rollout will take place over the coming weeks, while API customers globally can already opt in to text watermarking for selected models.

The technology, called textGrain, is designed to embed a statistical signal into the way an AI model selects words and tokens. The signal is intended to be invisible to readers and should not noticeably alter the meaning or style of the generated text. OpenAI says the system uses a secret key and a dedicated detector to determine whether the watermark is present, allowing authorized parties to verify the provenance of text without displaying any visible mark.

OpenAI is taking a cautious approach because text watermarking remains an emerging technology with limitations. The company says its detector will initially be available only to approved researchers and expert organizations rather than being publicly accessible. A watermark can provide evidence that text contains a signal associated with an OpenAI system, but it does not establish who wrote the text, whether the entire passage was generated by AI or whether the content itself is accurate.

The move also comes as other AI companies experiment with similar provenance technologies. Google DeepMind uses SynthID for AI-generated media, while Anthropic has introduced invisible marking for text produced by newer Claude models. These systems are intended to make AI-generated material easier to identify as AI becomes increasingly integrated into education, business, publishing and online communication. However, researchers have warned that text signals can be weakened or removed through extensive rewriting, translation and other transformations.

OpenAI says text provenance is only one part of its wider approach to identifying AI-generated content. The company already uses C2PA metadata and SynthID watermarking for supported images and audio, and it has developed verification tools for certain OpenAI-generated media. The expansion into text represents another step toward creating a broader system for tracking the origin of AI-generated material, although OpenAI acknowledges that no single watermarking or detection method is completely reliable.