OpenAI to Implement Invisible AI Watermark on EU ChatGPT Content

Image credit: Abondance

BRUSSELS — OpenAI will implement an invisible watermark, called textGrain, on content generated by its ChatGPT and Codex models within the European Union in the coming weeks to comply with the region’s new AI Act.

The watermark aims to embed a statistical signal into word choices, which a specialized detector can then identify, marking a step towards regulatory adherence for AI-generated content across all OpenAI plans in the EU.

OpenAI stated that the textGrain detector will not be publicly available at launch, with access restricted to approved expert researchers and organizations.

Internal tests conducted by OpenAI indicated that the watermark’s effectiveness could be significantly reduced by content modification. Replacing 25 percent of words in a 400-token passage with synonyms decreased detection rates from 92 percent to 17 percent.

The company also acknowledged that the watermark’s reliability diminishes in shorter texts, specifically those around 200 tokens, and in content types that inherently limit word choices, such as mathematical equations.

Globally, OpenAI has offered an API allowing clients to voluntarily activate the watermark on certain models since October 5, 2026, though it remains disabled by default for these international users.

OpenAI reported that the integration of textGrain does not substantially degrade performance across key benchmarks, including Artificial Analysis Intelligence Index, DeepSWE v1.1, Terminal-Bench 4.0, and GPQA Diamond.

The company clarified that a detected watermark only indicates that text was generated or processed by an OpenAI system. It does not confirm human contribution, ownership, responsibility, identity, or the accuracy of the content.

Source: Abondance