OpenAI Introduces Invisible Text Watermarking for ChatGPT and Codex in EU
OpenAI announced it is rolling out an invisible statistical text watermarking technology called 'textGrain' for ChatGPT and Codex text outputs in the European Union. The technology embeds subtle statistical signals during the word selection process to allow detectors to verify AI-generated text in compliance with EU AI Act rules. This represents one of the first major technical compliance measures deployed by a top AI company to satisfy the EU AI Act's transparency mandates. It provides a blueprint for how generative AI developers can prove text provenance without affecting readability or user experience. The watermarking feature will not be enabled globally by default, though API clients worldwide can opt in immediately to watermark their outputs. OpenAI is also granting detector tool access to approved researchers on a case-by-case basis, ensuring the tool detects watermarks without exposing user identities or prompt content.
## BACKGROUND
Statistical text watermarking embeds subtle mathematical signals into a large language model's token sampling process without degrading the fluency or meaning of the generated text. Under the EU AI Act, providers of generative AI systems must ensure that AI-generated content is marked in a machine-readable format to mitigate risks like misinformation and academic dishonesty.