OpenAI rolls out invisible text watermarks in ChatGPT and Codex for the EU

Invisible watermarking lands in ChatGPT and Codex for the EU

OpenAI is adding an invisible, machine-readable watermark to eligible text produced by ChatGPT and Codex in the EU, embedding the pattern in the wording itself rather than in metadata. The rollout is a compliance move tied to the EU AI Act, which requires model-made text to carry a machine-readable identifier that flags it as AI-generated. The company is positioning the marks as provenance signals rather than hard proof: OpenAI states that watermarks are often undetectable, especially in short passages, and that rewriting or translating text can completely remove the watermark.

The community thread also notes the watermark "will remain off by default in the API," which means API builders shipping text into the EU will not see it unless they opt in.

Where the watermark applies and where it does not

Coverage converges on the same scope: ChatGPT and Codex outputs in the EU receive the watermark, with detection framed as starting with researchers rather than being exposed as a consumer flag. Gizmodo frames the rule as requiring "a machine-readable identifier that flags the" model-made text, while the community thread clarifies the API remains opt-in for the marks. Editors and platform teams shipping EU-facing chat or code assistants should expect the watermark to be on by default in the consumer products and off in the API, and should plan for the fact that detection is probabilistic, not absolute.

Practical limits editors and engineers need to plan around

TechCrunch reports that editing can make the invisible marks harder to detect, and OpenAI's own account extends the warning to translation and short passages. For practitioners, that means three failure modes are explicit at launch: editing the output, translating it, and truncating it can all weaken the watermark below a reliable detection threshold. Teams building detection dashboards or provenance audits into editorial workflows should treat watermark scores as a weak signal rather than a binary verdict, and should not rely on the mark alone for compliance reporting on translated or lightly edited EU content.

Researcher access as the first detection channel

OpenAI's own post frames access to detection as starting with researchers, signaling that the verification side of the system is being rolled out in a controlled order. That order matters for anyone building detection tooling or for newsroom standards desks evaluating whether to require AI-text provenance: the consumer-facing detector is not the first thing being shipped, the research API is. Practitioners evaluating detectors should expect a staged onboarding and should design their provenance pipelines to ingest watermark scores alongside other signals, not as the sole input.

What to verify next

The clearest concrete checks a reader can run on 2026-10-06 are: whether their ChatGPT workspace is in the EU rollout cohort, whether their Codex integration is consumer or API (API keeps the watermark off by default), and whether any in-house detector is ready to consume watermark scores when OpenAI opens researcher access. No evidence line in this set names a release date for the detection endpoint or a deadline for full EU coverage, so any forward-looking date for detection rollout is intentionally left qualitative here.

For teams handling multi-step text pipelines, the open question worth tracking is how robust the watermark remains after common editing passes such as translation and rewriting.