OpenAI has announced the introduction of textGrain in ChatGPT and Codex. This technology functions as an imperceptible watermark in texts, aiding in the identification of content produced by artificial intelligence.
Initially, this feature is only available to users located in the European Union. Although OpenAI had previously announced a watermarking method in 2024, it had not been adopted on a large scale until now.
The decision to implement textGrain was motivated by the need to comply with the European Union's Artificial Intelligence Act (EU AI Act). This regulation requires entities providing generative AI services—covering text, image, audio, or video—to incorporate mechanisms to label content as synthetic through verification tools.
While a similar feature already existed for media content in ChatGPT, the implementation for texts occurs now. It is important to note that the invisible watermark is not a legal requirement, but rather the method chosen by OpenAI to comply with European legislation.
OpenAI itself defines textGrain as an 'invisible statistical signal.' This implies that the generated material does not receive hidden characters or extra spacing, but rather discrete adjustments that statistically modify how the AI model selects words in the text, all without compromising the final quality of the result.
In the coming days, this technology will be applied by default to all texts created by ChatGPT and Codex for all user plans based in the European Union.
Additionally, even with the regional restriction in this initial phase, developers globally can enable the watermark in certain OpenAI API models for use in their own applications.
However, OpenAI issues warnings about the system's limitations: the watermark may be weakened through edits or certain types of text, and its accuracy may be compromised in short content. The organization also clarifies that textGrain does not identify who requested the text, cannot attest to the veracity or falsehood of the result, nor quantify human contribution in modified content.