OpenAI Adds Hidden Watermark to ChatGPT Text in EU, but Rewriting 20% May Defeat It

OpenAI is beginning to add an automatic watermark to text generated by ChatGPT in the European Union, in an effort to help identify content written by artificial intelligence.

The hidden marker is intended to provide a technical way to distinguish AI-generated writing from human-written material. Unlike a visible label, it is described as being embedded within the text itself.

The approach may have a significant limitation: changing 20% of the wording could be enough to defeat the watermark. The available information does not explain how that threshold was tested or which kinds of edits could make the marker ineffective.

That potential weakness highlights a challenge for AI-content detection. A marker may help identify an original output, but its usefulness could diminish once the text has been rewritten.