AIBRIEF
OpenAI's new 'textGrain' watermark marks AI generated text in ChatGPT and Codex under EU AI Act rules. OpenAI itself warns the watermark cannot prove truth, authorship, or ownership.
ChatGPT and Codex text generated in the EU will carry an invisible watermark flagging it as AI-generated, OpenAI's first deployment of a system it calls textGrain.
The watermark embeds a machine-readable signal in token patterns rather than visible text, letting a separate detector flag a passage as likely AI-generated without reading it. OpenAI says the system "matched or exceeded" Google DeepMind's SynthID on its benchmark suite, with similar quality scores between watermarked and unwatermarked outputs.
The rollout is regional at launch. Eligible ChatGPT and Codex users across all EU plans see watermarked text; users elsewhere do not. OpenAI's announcement frames the move as compliance with the EU AI Act, the bloc's risk-based AI regulation that has been phasing in since 2024 and now treats provenance signals as a core requirement for general-purpose AI. API customers worldwide can opt in to watermarked outputs from select models today.
The honest limits are in the technical report: textGrain "does not guarantee reliable detection," and a watermark "does not verify the accuracy of the underlying text, determine ownership, measure human contribution, or prove human authorship." The detector itself is not public; approved researchers and expert organizations can apply for case-by-case access under the EU Code of Practice. They learn whether a string carries an OpenAI watermark, not who wrote it or why.
Anthropic shipped a similar rollout in August 2026, also citing the AI Act, with SynthID at its core. The pattern is becoming a vendor baseline, not a one-off.