OpenAI will start hiding a watermark in ChatGPT text for users in the European Union. The mark is invisible, and it’s built to survive copying and reformatting. It could make AI-written text easier to spot. It also has limits creators should know about.

The move is part of the EU’s effort to enforce its AI Act, which requires transparency on AI-generated content. OpenAI’s answer is called textGrain. It embeds the watermark as the model writes.

How OpenAI’s text watermark works

Watermarks usually sit on top of a finished file, like a logo on an image. TextGrain is different because it’s built into the way the language model picks words.

Advertisement

ChatGPT will lean toward some words over others in patterns no reader would notice. Those patterns add up to a statistical signal. Software built to look for the signal can then flag the text as AI-generated.

OpenAI wants the signal to hold up when someone copies, pastes or reformats the text. That durability leaves some creators feeling uneasy.

What AI watermarks can’t tell us

The detector can say whether AI was involved. It can’t say how much of the text came from a person and how much came from the model. Even the yes-or-no answer has holes, and there are a few important limitations noted by OpenAI in its Oct. 5 announcement.

Editing weakens the signal fast. In OpenAI’s tests on 400-token passages, detection started at about 92%. Replace 10% of the words with synonyms and it fell to about 66%. At 25%, it dropped to about 17%.

Partial use counts too. Say you write a blog post, then ask ChatGPT to rewrite a few paragraphs or polish a sentence or two. That text may carry the watermark. Right now, nothing separates that scenario from content written entirely by AI.

Short or tightly constrained text is harder for the detectors to read. A missing watermark proves nothing either. OpenAI says so itself.

What we think

For video creators, the risk is small but real. Scripts and show notes often start in a chat window. A ChatGPT-assisted script could carry the mark even after you rewrite parts of it.

We’d treat the watermark as a weak signal, not a verdict. Light editing wrecks detection, and the public can’t run the detector anyway. Creators who say up front when they use AI shouldn’t lose sleep over it.

Pricing and availability

OpenAI hasn’t announced a separate fee for textGrain. The watermark will be added to eligible ChatGPT and Codex text in the EU across all plans over the coming weeks. API users worldwide can opt in on select models starting Oct. 5, but the feature is off by default.

The detector isn’t public yet. OpenAI is granting access to qualified researchers and organizations first, who will work with the company on improvements. There is still much work to do.