OpenAI is adding watermarks to ChatGPT text in the EU: Here’s what it means

OpenAI is set to introduce an invisible watermark to text produced by ChatGPT and Codex, in an effort to adhere to the EU AI Act's transparency guidelines. This watermark will be rolled out in the coming weeks to all ChatGPT and Codex users in the EU, while currently limited to eligible users on all plans. Developers worldwide can opt to turn on the feature for select models.

The watermark, named textGrain, adds an "invisible statistical signal" to the model's word choices, which can then be detected by a scanner. OpenAI claims that textGrain matches or surpasses other watermarking technologies, like Google's SythID. However, the company acknowledges that the technology is still in its early stages and has limitations, as reliable detection is challenging in shorter or edited text.

While the watermark does not determine ownership or responsibility, nor identify the user, it does not imply human authorship or guarantee accuracy. Article 50 of the AI Act, which comes into effect on August 2, 2026, mandates transparency obligations for AI providers and deployers, including the addition of machine-readable marks to detect AI-generated or manipulated content.

OpenAI's move aligns with Anthropic's decision to add invisible watermarks to Claude models to comply with the EU AI Act's transparency requirements.