OpenAI rolls out invisible textGrain watermarks for ChatGPT and Codex under EU AI Act
The European Union's transparency regime for artificial intelligence has been in force since August 2, 2026, when Article 50 of the bloc's AI Act began to apply to providers of systems that generate synthetic audio, image, video or text. Under those rules, set out in Regulation (EU) 2024/1689, such providers must ensure that outputs are marked in a machine readable format and are detectable as…
The European Union's AI transparency rules, enforced since August 2, 2026, require providers of AI systems that generate synthetic content to mark outputs in a machine-readable format and make them detectable as artificially created. Deployers must also disclose AI-generated text published for public interest, though human-reviewed content is exempt. OpenAI responded to these regulations by introducing an invisible watermark for text generated by ChatGPT and Codex within the EU, with rollout planned for the coming weeks.
The watermarking system, called textGrain, subtly influences the model's word choices, embedding a statistical pattern in the text that can be detected with a detector holding a secret key. The watermark is not visible but persists in copied and pasted text, ensuring it travels with the content. OpenAI is offering the watermarking feature to API customers worldwide, allowing select models to be enabled from the launch day.
The company is collaborating with cloud partners to extend watermarking to model outputs accessed through their services.
OpenAI has invited only approved researchers and expert organizations to access the text watermark detector initially, emphasizing the importance of a vetted group in verifying the technology. Detection rates are presented as a target of 1 percent false positives, with successful detection in 80 percent of 200-token passages and 95 percent of 400-token passages for content like psychology.
However, editing text significantly impacts detection, reducing rates from 92 percent to 17 percent when 10 percent and 25 percent of words are replaced with synonyms, respectively.
Despite OpenAI's claim that the watermark does not reveal human contribution, ownership, or verify accuracy, the technology aims to shift the default, making detected hits meaningful evidence. OpenAI plans to release the watermarking technology as open source, encouraging outside scrutiny of the false positive rate and promoting transparency.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
