OpenAI is adding an invisible watermark to ChatGPT text

OpenAI has announced that it will add an invisible watermark to text generated by ChatGPT and Codex in the European Union. The reason is the transparency rules of the EU AI Act, which require AI-generated content to be marked in a way other systems can recognise.

How does a text watermark work?

The watermark is not a symbol or a hidden character. OpenAI's method, called textGrain, subtly steers the model's next-word choices with a secret key; hundreds of small nudges form a pattern that readers cannot see but a detector can pick up. Because the pattern lives in the words themselves, it travels when text is copied and pasted. According to the company, the watermark does not identify the user and caused no meaningful change in model performance.

Who is affected, when and where?

In ChatGPT and Codex the watermark will reach eligible users on all plans, in the EU only, over the coming weeks. Developers using the API anywhere in the world can switch it on for select models from today; it is off by default. According to Search Engine Journal, the transparency rules began to apply on 2 August, and systems placed on the market before that date have until 2 December to meet the marking obligation.

The detector is not public for now; it is available only to approved researchers and expert organisations.

Does editing remove the watermark?

It weakens it. In OpenAI's test on 400-token English passages, replacing 10 per cent of words with synonyms cut detection from about 92 per cent to 66 per cent; Search Engine Journal reports that at 25 per cent the rate fell to 17 per cent. Detection is harder still for short passages, maths answers and translated text. OpenAI also says a missing watermark does not prove human authorship: the text could be too short or heavily edited, or it could come from another company's AI.

What should content teams in Turkey do?

The rollout is limited to the EU for now, so a team or agency in Turkey using ChatGPT will not see the watermark by default. Still, for brands with customers in the EU and for teams producing content through the API, three steps make sense: record which content was produced with which tool; confirm with your provider whether the watermark is switched on for your API use; and base quality control on editorial review and source checking rather than on watermark detection. A watermark's absence and its presence say nothing about whether a text is accurate.