OpenAI turned on an API opt-in for text watermarking on October 5, 2026. It is off by default. Eligible ChatGPT and Codex text generated in the EU will carry an invisible watermark over the coming weeks, and OpenAI says applications for detector access are now open.
The timing matters because EU provenance rules are pushing providers to prove where text came from, and OpenAI is answering with a mark embedded in output rather than a label stapled onto a response. That distinction is the whole story. An invisible watermark has to survive copying, formatting, translation, and whatever your pipeline does to text before a human sees it. Nothing in the announcement says how durable the mark is against those steps, and OpenAI has not disclosed detection accuracy, false positive rates, or which model versions carry the feature.
For teams running API integrations against OpenAI models, the opt-in default is the operational fact to act on now. If a client, a compliance team, or a downstream partner expects watermarked output, someone has to flip that setting, because nothing changes automatically. Checking whether EU-region traffic is even separable in your own routing logic is a reasonable first step, since the watermark only applies there. Anyone building moderation, plagiarism checks, or audit logging on top of OpenAI text should confirm whether their current pipeline strips formatting in ways that could interact with an invisible mark, though OpenAI has not said what breaks it.
Detector access is the second lever, and it is gated. OpenAI says applications are open, but the announcement does not specify eligibility criteria, turnaround time, or whether detection will be offered as an API endpoint versus a manual review process. Platform operators who need to verify content at scale, rather than one document at a time, should treat this as unresolved until OpenAI publishes mechanics. A detector that requires an application and a wait is not a service you can wire into an automated content pipeline today.
On-call engineers should not expect this to show up as an incident, but it can show up as a support ticket. If a customer asks why their EU text output is unmarked, the answer is the default: nobody turned it on. If a customer asks why text outside the EU is unmarked, the answer is scope: the rollout is EU-specific. Worth checking now whether your status dashboards or client docs reference watermarking at all, since a feature that is off by default and silent in behavior will not surface in error logs or alerts.
What is confirmed is narrow: an API opt-in exists, it started October 5, 2026, it targets EU text from ChatGPT and Codex, and it is off unless someone turns it on. What is not confirmed includes detection accuracy, durability of the watermark through common text transformations, and the shape of detector access once applications are reviewed. Teams that depend on provenance claims for compliance reporting should flag the gap in writing now, before an auditor asks whether the watermark survived a copy-paste into a different document.