Back to News
NewsRegulationAi

OpenAI to Introduce Text Watermarking in the EU

OpenAI is launching text watermarking in the EU to comply with the AI Act, embedding invisible marks in responses from ChatGPT and Codex.

Mercation Admin
October 6, 2026 2 min read
OpenAI to Introduce Text Watermarking in the EU

OpenAI plans to introduce machine-readable text watermarking to comply with the EU AI Act, with the feature rolling out in the coming weeks. This initiative will see invisible watermarks embedded in responses from ChatGPT and Codex for EU users.

We’re expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. This work builds on our existing tools that help verify whether an image or audio file was produced with our models.

— OpenAI (@OpenAI) October 5, 2026

Despite this advancement, OpenAI admits that the current technology cannot reliably trace the origin of any text.

Details of the Watermarking Process

The upcoming watermark will be applied to eligible text from ChatGPT and Codex, available across all subscription plans. The technology, called textGrain, uses a statistical signal based on the model's word choice, which can be detected by a specialized tool.

This effort is a response to the EU AI Act, which requires generative AI providers to make generated text identifiable. OpenAI has acknowledged that current watermarking methods are still in early development.

Challenges in Detection

The effectiveness of textGrain varies based on the text's length and content. Tests show a 1% false positive rate, with successful detection in about 80% of 200-token texts and around 95% in 400-token texts. However, detection rates drop significantly for mathematical texts due to less variability in language.

Editing also complicates detection; for instance, replacing 10% of words with synonyms decreased the detection rate from 92% to 66%, and a 25% replacement dropped it to 17%.

Because of these challenges, OpenAI is not making the detection tool publicly available yet but is inviting researchers to help evaluate its reliability.

Limitations of Watermarking

OpenAI cautions that a watermark does not definitively indicate AI authorship. It does not convey the extent of human involvement or legality and does not identify the user or their interactions. Conversely, a lack of a watermark does not mean the text was created by a human, as the signal can be lost through editing or if the text is too brief.

Optional Watermarking for API Users

OpenAI is also introducing an optional text watermarking feature for API users worldwide, available for specific models and off by default. The company intends to release the technology as open-source in the future to allow for wider integration.

Initially, OpenAI will focus on testing the watermarking system in the EU to collect performance data before considering a broader rollout.

Broader Content Provenance Efforts

This watermarking initiative is part of OpenAI's larger strategy to enhance content provenance systems. The company already uses Content Credentials and C2PA standards for images, along with invisible watermarks called SynthID. OpenAI believes that a multi-faceted approach is necessary, as metadata can be easily removed, and text can be rewritten or translated.