In plain English
- What happened
- OpenAI said that beginning October 5, 2026, API customers worldwide can opt in to add an invisible watermark to AI-generated text, and that it will add the same watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks. The move responds to the EU AI Act's requirement that generative AI providers make generated text identifiable in a machine-readable way.
- Why it matters to your business
- Any business serving EU customers through a product built on OpenAI's API, or using ChatGPT inside the EU, may see this watermark appear in outputs without asking for it. Businesses outside the EU can also turn it on voluntarily if they want a transparency signal, but OpenAI's own data shows the watermark is easy to erase with light editing, so it is not a tool to rely on for proving what is or is not AI-written.
- Do this
- Check your OpenAI API dashboard for the new watermarking setting if your business builds a product on top of it, and read the help center article OpenAI links from this announcement before deciding whether to turn it on.
The watermark itself is a feature called textGrain. OpenAI describes it as an invisible statistical signal woven into the model's word choices, which a separate detector tool can later check for. The detector is not public yet.
What's rolling out, and where
Watermarking stays off by default in the API everywhere; customers have to turn it on. Inside the EU, OpenAI plans to add the invisible watermark to eligible ChatGPT and Codex text output across all plans over the coming weeks, without an opt-out mentioned for end users there.
Access to the detector that reads the watermark is limited at launch. OpenAI is opening applications to researchers and expert organizations, who will be approved case by case under the EU's Code of Practice, rather than making the checker available to the public or to businesses that want to verify content themselves.
How reliable the watermark actually is
OpenAI published its own detection numbers alongside the announcement, and they come with real limits. Shorter passages are much harder to catch than longer ones, and content with less flexibility in wording, like math, is harder to detect than content with more room for phrasing, like psychology answers.
| Scenario | Detection rate |
|---|---|
| 200-token passage (psychology) | ~80% |
| 400-token passage (psychology) | ~95% |
| 10% of words replaced with synonyms | 66% (down from 92%) |
| 25% of words replaced with synonyms | 17% (down from 92%) |
Replacing just 25% of the words in a passage dropped OpenAI's own detection rate from 92% to 17%, according to the company's published evaluation.
On the benchmarks OpenAI uses to test its latest model, the company reported no meaningful quality drop between watermarked and unwatermarked text, so turning it on is not expected to change how an answer reads.
What the watermark does not prove
OpenAI is explicit that a detection result does not establish ownership, accuracy, or who wrote something. It can indicate that an OpenAI system touched a passage, but it says nothing about how much human editing went into it, and it does not identify a specific person, account, or conversation.
Reality check
A watermark is not a way to prove a piece of writing was or was not made with AI, and treating it as proof of authorship in a dispute with a client, a student, or a contractor would be a mistake. OpenAI itself says the absence of a detected watermark does not mean a human wrote something; the text may simply be too short, too edited, translated, or from a different model entirely.
- 1
Check whether you serve EU customers
If any product you run touches users in the European Union through ChatGPT or the API, expect the watermark to begin appearing in text output there over the coming weeks, even if you did not request it.
- 2
Decide whether the API opt-in is worth turning on
If you build your own product on top of OpenAI's API, the watermark is off by default; weigh whether a transparency signal is useful for your customers against the fact it adds no visible marker and can be stripped by light editing.
- 3
Do not treat the watermark as a compliance shield
The EU AI Act requirement is about machine-readable identification of generated text, not about proving human authorship or settling disputes, and OpenAI's own numbers show the signal is weak once text is edited.
Sources
