OpenAI textGrain watermarks EU ChatGPT text, API is opt-in

OpenAI is adding an invisible watermark to ChatGPT and Codex text in the European Union, and it opened the same textGrain watermark to API customers worldwide today. In the EU it switches on automatically in the coming weeks. In the API it is off by default, and you decide per project. If you ship AI-written copy, emails, or code to customers, you now have a setting to own and a policy to write.

What actually happened

OpenAI published its approach to EU text provenance rules on October 5. The watermark, called textGrain, nudges which words the model picks when several are near-equal ("fast" vs. "quick"). Those choices are tied to a secret key, so a detector holding the key can spot the statistical pattern across a passage. A reader sees nothing. OpenAI released a technical report on the method.

The rollout, per TechCrunch and BleepingComputer:

• EU ChatGPT and Codex: eligible text output gets the watermark automatically "over the coming weeks," tied to the EU AI Act transparency rules.
• API, worldwide: opt-in for select models starting today. You turn it on per project in Settings, or for the whole org, and pick which models carry it.
• Detector: applications are open, but access starts with approved researchers and expert organizations. You cannot check your own text yet.

OpenAI's own numbers show the limits. Detection runs about 80% on 200-token responses and about 95% on 400 tokens at a 1% false-positive rate. Swap 10% of words for synonyms and detection drops from roughly 92% to 66%. Swap 25% and it falls to 17%. Math and translated text are harder to catch. OpenAI says plainly that a missing watermark does not prove a human wrote it, and the mark does not identify the user or account.

Why the OpenAI watermark matters for your business

Two vendors, two defaults. Anthropic watermarks Claude text globally with no opt-out. OpenAI keeps the API off unless you flip it. If you route between providers, your output is now marked or unmarked depending on which model answered. Decide what you want and make it a config value, not an accident of routing.

Don't build a "detect AI" feature on this. At 17% detection after light edits, any workflow that grades customer, applicant, or vendor text on watermark presence will produce wrong answers. Treat it as a weak provenance signal, not a lie detector.

EU-facing teams: document it. If you serve EU customers with generated content, the AI Act transparency duty is already live. A written note of which models mark output, and where, is cheap evidence when someone asks.

Sources: OpenAI, OpenAI textGrain technical report, TechCrunch, BleepingComputer.