OpenAI to Watermark ChatGPT Text by Default in EU as AI Act Bites in 2026
Updated
Updated · Ars Technica · Oct 6
OpenAI to Watermark ChatGPT Text by Default in EU as AI Act Bites in 2026
3 articles · Updated · Ars Technica · Oct 6
Summary
OpenAI said ChatGPT text generated in the European Union will now carry an automatic watermark, while the same feature will remain optional and off by default in other regions.
The shift is aimed at complying with the EU AI Act, in force since August, which requires AI-generated content to be marked in a machine-detectable way.
OpenAI’s proprietary system, called textGrain, embeds subtle patterns in word choice that specialized tools can detect without noticeably changing output quality.
Detector access will be limited at first to approved researchers and organizations, with a request process for others, even as OpenAI acknowledges no watermarking method is fully reliable or hard to evade.