EU Gains ChatGPT Watermark as OpenAI Keeps Detector Ready

OpenAI is introducing a hidden ChatGPT watermark across chatbot text and Codex coding output in the EU. But company testing found that replacing one in four words with synonyms cuts detection to only 17%.

The step is a response to the EU AI Act, the bloc's rules for artificial intelligence, which require machine-readable markers on AI text. OpenAI acknowledges the approach has not matured, and is therefore keeping the detector private for the time being.

How Much Editing Can the ChatGPT Watermark Withstand?

We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology.

Our tools already help verify whether an image or audio file was created with our…

— OpenAI (@OpenAI) October 5, 2026

The system is dubbed textGrain, per the announcement, and it steers the model's wording selections to leave a concealed statistical footprint. A standalone detection tool then checks passages for the pattern.

OpenAI's FAQ says the marker sits in the text's wording alone. No hidden characters, invisible spaces, or unusual punctuation are added. As a result, the usual cleanup trick of stripping out hidden characters finds nothing to delete.

Size has a big impact. OpenAI counts text in tokens - the word pieces that AI models handle. A token represents roughly three-quarters of an English word.

Testing by OpenAI showed the detector flagged roughly 66% of unedited responses around 150 words in length. With approximately 300 words, the detection rate climbed to about 92%.

Even modest edits hurt the ChatGPT watermark. In 300-word passages, substituting one word out of ten with a synonym brought detection down from 92% to 66%. Changing every fourth word dropped it to just 17%.

The marker left benchmark results for GPT-6 Astra, OpenAI's newest frontier model, virtually unchanged.

Why Keep the Detector Out of Public Hands?

OpenAI is not releasing a public tool; instead it is accepting applications from vetted researchers and specialist organizations. A positive match exposes neither the user, the prompt, nor the conversation, according to OpenAI.

A negative outcome also proves little. Brief, modified, or translated passages can evade the detector, as can material generated by competing AI systems.

The absence of a detected watermark does not prove human authorship.

Source: OpenAI

Visual and audio content follows another path. Users can upload such files to OpenAI's public verification tool and look for the SynthID marker.

The ChatGPT watermark defaults to off outside Europe. API clients everywhere, though, can now enable it for certain models through project or organization settings.

The introduction arrives after the EU began enforcing AI Act transparency rules in August. In crypto, the impact is tangible: TRM Labs recorded a 40% year-over-year rise in criminal AI use.

So the divide between regulatory requirements and what detection tools can actually prove will probably shape compliance strategy at every major AI lab.