TL;DR

OpenAI set out on Monday 5 October how it will meet the EU AI Act’s requirement that AI-generated text be machine-identifiable. In the next few weeks, text produced by ChatGPT and Codex for EU users will start carrying a hidden watermark. API customers anywhere can switch watermarking on from now, and access to the detector is restricted to approved researchers because OpenAI says the risk of false positives and missed watermarks is too high for public release.

How the rollout works

The watermark, called textGrain, works by nudging the model’s choice of words to leave a statistical pattern a detector can pick up. OpenAI says it performed as well as or better than the alternatives it tested, including SynthID for text, and that it plans to release the technology as open source.

There are three parts to the launch:

• EU consumers: every ChatGPT and Codex plan is covered where eligible, but only inside the EU.
• API customers: watermarking is off by default but can be enabled for selected models, wherever the customer is based.
• Detection: researchers and expert bodies can apply for access, granted case by case. The tool says only whether a watermark is present, not who wrote the text.

The limits OpenAI admits

The company’s own figures show why it is cautious. Holding false positives at 1%, the detector caught roughly 80% of watermarked passages of 200 tokens and around 95% at 400 tokens, for flexible subjects such as psychology. Mathematical text, which leaves less room in word choice, fared much worse. Light editing does real damage: swapping 10% of words in a 400-token passage for synonyms dropped the detection rate to 66%, from about 92%, and swapping a quarter cut it to 17%.

OpenAI is also explicit that a watermark does not show how much a human contributed, does not identify the user, and says nothing about accuracy. Failing to find one proves nothing about human authorship. It reports no meaningful benchmark difference between watermarked and unwatermarked output from Astra, its latest model.

Looking forward

For UK organisations the effect is indirect but real. The ChatGPT watermark is limited to the EU, so British users are outside it, but a UK business generating text through OpenAI’s API for European customers now has an opt-in it can weigh against its own transparency obligations, which is how OpenAI frames the choice. In our view, the detection numbers matter more than the rollout: a signal that fades after modest rewording is a weak basis for anyone hoping to use it to police AI-written coursework, job applications or news. OpenAI itself says no single provenance technique is enough by itself.