## OpenAI: invisible text watermarking for ChatGPT and Codex — EU-first

In a 3-post thread, OpenAI said it is expanding content provenance from image/audio to text, "in response to EU regulatory requirements":

In the EU, eligible text from ChatGPT and Codex will be watermarked over the coming weeks to comply with the EU AI Act.
API customers worldwide can turn on text watermarking for select models starting today — opt-in, not a global default at launch.
OpenAI describes it as a phased, regional approach to leave room to learn: *"We expect to revisit each part of this approach as the technology, standards, and evidence evolve"* — quoted from the linked EU text provenance page.

Answers one question only: *"Was this likely generated by an OpenAI model?"*
Embeds an invisible statistical signal in wording as generated — no visible marks or special characters.
Does not say who wrote it or who owns it; does not identify a person, organization, account, conversation or prompt.
In OpenAI's testing, it hasn't affected capability, speed, or how responses read.

Often undetectable, especially in short passages.
Rewriting or translating can completely remove the watermark.
Therefore only approved researchers get the detector for now, to help evaluate and improve the technology. OpenAI calls text watermarking ongoing research with feedback from users, developers, policymakers and researchers.

@kimmonismus notes the watermark is embedded in word choice itself, criticizes detection requirements influencing phrasing, and cites a test where replacing 25% of words with synonyms on 400-token passages cut detection from ~92% to ~17%. Also notes it cannot distinguish human contribution — text you wrote and had ChatGPT edit can carry it.
@ns123abc highlights: *"We are not making text watermarking a global default AT LAUNCH"* — phased starting with the EU.

No cutoff, model list for API opt-in, or detector-access details beyond "approved researchers" is established in the X thread data.