OpenAI Adds Invisible Watermarks to AI-Generated Text in ChatGPT & Codex
Staff Reporter
07 October 2026, 15:09
OpenAI is introducing an invisible watermarking system called TextGrain to help identify AI-generated text created with ChatGPT and Codex. The technology is initially being introduced for users in the European Union to comply with the EU AI Act’s transparency requirements. Instead of inserting hidden characters or extra words, TextGrain creates a statistical signal by subtly influencing how the AI model selects words.
OpenAI says its tests detected the watermark in around 80% of 200-word texts and about 95% of 400-word texts. However, editing, translation, short texts and other AI systems can make detection less reliable.
In this video, we explain how OpenAI’s invisible AI watermark works, how accurate it is, its limitations, and what it could mean for identifying AI-generated content.
