Our approach to EU text provenance rules
OpenAI is rolling out invisible text watermarks called textGrain for ChatGPT and Codex in the EU to comply with the EU AI Act, while keeping them optional for global API customers.
summary is enough The announcement outlines compliance mechanics and technical limitations, which the skeptical comment thread contextualizes well.
summary
OpenAI is introducing textGrain, an invisible statistical watermark, to comply with the EU AI Act for EU users while offering it as an opt-in for global API customers. The company acknowledges significant technological limitations, noting that detection rates drop sharply with short texts, mathematical content, or simple synonym substitutions. Because of high false-positive and evasion risks, detector access is initially restricted to approved researchers. Commenters widely dismiss the implementation as ineffective regulatory theater that clutters outputs and is trivial to bypass.
key points
- The textGrain watermark embeds an invisible statistical signal by biasing token probabilities during model sampling.
- Detection accuracy drops severely with text length, strict domain constraints, and basic editing (replacing 25% of words drops detection to 17%).
- Watermarking rolls out automatically for EU ChatGPT and Codex users, but stays optional for global API customers.
- Detector access is strictly limited to approved researchers due to high false-positive risks and easy circumvention.
who should read it
Engineers and product builders navigating AI compliance and the practical enforcement of the EU AI Act.
discussion skeptical
consensus
Commenters widely view text watermarking as security theater that is easily bypassed via paraphrasing, translation, or synonym replacement, while polluting code and text outputs with unnecessary friction.
pushback
A minority argues that even easily bypassed regulations establish legal definitions of intent, making it easier to prove deception in court if bad actors intentionally strip watermarks.
Worth noting that this is what OpenAI wrote about text watermarking two years ago:
> While it has been highly accurate and even effective against localized tampering, such as paraphrasing, it is less robust against globalized tampering; like using translation systems, rewording with another generative model, or asking the model to insert a special character in between every word and then deleting that character - making it trivial to circumvention by bad actors.
> Another important risk we are weighing is that our research suggests the text watermarking method has the potential to …
It highlights OpenAI's own past warnings that text watermarking is easily gamed and risks stigmatizing non-native English speakers.
Generated by an LLM from the article and thread. Check the sources before relying on it.