OpenAI will begin adding invisible watermarks to AI-generated text from ChatGPT and Codex in the European Union as the company prepares to comply with transparency requirements under the EU AI Act, according to TechCrunch.
The technology, called textGrain, is designed to embed a hidden statistical pattern into generated text by subtly influencing word choices. Unlike a visible label, the watermark remains embedded when users copy and paste the text.
How the Watermark Works
Rather than adding visible characters or metadata, textGrain adjusts the selection of words during generation to create a pattern that can later be detected.
The approach is intended to help identify content produced by OpenAI’s systems while preserving the readability of the original text.
The rollout will cover both ChatGPT and Codex in the European Union.
Detection Can Weaken When Text Is Modified
Testing cited in the reports also points to limitations in the technology.
OpenAI’s tests found that replacing about 10% of the words in watermarked text with synonyms reduced detection rates from approximately 92% to 66%.
That result illustrates one of the challenges facing text watermarking systems: unlike images or other digital files, written content can be easily edited, translated, summarized or rewritten after generation.
The introduction of textGrain comes as AI companies prepare for European rules requiring greater transparency around certain AI-generated or manipulated content.
