OpenAI Begins Phased Text Watermarking Under the EU AI Act

OpenAI has begun rolling out text watermarking as part of its response to Article 50 of the EU AI Act. API customers worldwide can now opt in to watermarking for select models, while eligible ChatGPT and Codex output in the European Union is expected to receive invisible watermarks over the coming weeks. The feature is not being enabled globally by default at launch.

The technology, called textGrain, embeds an invisible statistical signal into the word choices made during generation. A separate detector can then analyze a passage for that signal. OpenAI says it plans to open-source the technology and has begun offering detector access to approved researchers and expert organizations rather than releasing it publicly.

The company reports strong detection rates on longer, minimally edited passages, but the system has clear limits. In OpenAI’s testing, synonym replacement significantly reduced detection accuracy, and performance varies by subject matter depending on how much flexibility the model has in its wording. OpenAI also stresses that a detected watermark does not establish authorship, ownership, accuracy, or responsibility, while the absence of a watermark does not prove that text was written by a human.

The rollout reflects the EU AI Act’s requirement that providers make AI-generated text identifiable in a machine-readable form. OpenAI describes text watermarking as an early technology and says it will continue studying how well the signal survives editing and translation while exploring ways to distinguish AI assistance from AI authorship.

Footnote

1. Sophie Denar, OpenAI Begins Phased Text Watermarking Under EU AI Act Rules, Unite.AI, October 5, 2026: https://www.unite.ai/openai-begins-phased-text-watermarking-under-eu-ai-act-rules/