OpenAI introduces textGrain, the invisible watermark for AI-generated text.

OpenAI has announced the introduction of textGrain, a feature that applies an invisible watermark to texts generated by its artificial intelligence systems like ChatGPT and Codex. The update will be implemented in the coming weeks within the European Union, while outside the EU, it is already optionally available via API. The technology embeds a recognizable statistical pattern into the text without adding symbols or hidden characters.

The purpose is to make AI-generated text identifiable, aligning with the transparency rules of the AI Act that came into effect on August 2. However, OpenAI itself clarifies that this is not an infallible "truth machine": internal tests show a false positive rate of 1%, and the detection capability drops drastically if the text is modified, for example, by replacing 10% of the words with synonyms.

The European Union's AI Act, which came into force on August 2, 2026, has introduced new transparency obligations for companies developing artificial intelligence. This legislation represents the world's first comprehensive legal framework to regulate AI and requires, among other things, that automatically generated content be clearly identifiable. The goal is to combat misinformation and ensure that users know when they are interacting with content produced by machines.

OpenAI, the company behind ChatGPT, is complying with these rules following the example of other companies like Anthropic. The technical challenge is complex: creating a watermarking system that is reliable but not intrusive. Previous attempts have faced the difficulty of distinguishing human texts from AI-generated ones, especially when they are short or modified. The textGrain solution arises from this regulatory need and the growing concern over the indiscriminate use of chatbots in producing textual content, a phenomenon that exploded with the launch of ChatGPT in 2022.