OpenAI Adds Subtle Watermark to ChatGPT Text in Europe

OpenAI has begun adding an invisible statistical watermark called "TextGrain" to text generated by "ChatGPT" and "Codex" models in the European Union over the coming weeks, according to a report by platform "36Kr". The measure applies to all subscription types to meet the European AI Act requirement that generated text be machine-identifiable.

The watermark is added during word selection without altering the meaning or appearance, and the company plans to make it open source. API customers worldwide can enable it for certain models, while the default option remains disabled outside Europe, and detection tools are restricted to certified researchers and institutions.

Company tests indicate that detection accuracy reaches about 80% in texts of 200 tokens, rising to about 95% in texts of 400 tokens, with performance declining in mathematical content or when words are replaced with synonyms. The company confirms that adding the watermark makes no noticeable difference in performance.

In the same context, Thibault Sottiaux, head of the "Codex" division, announced a commitment to daily improvements over 28 days, most notably increasing the default speed for "GPT-6 Astra" and "GPT-6.1 Sol" subscribers from 30 to 50 tokens per second, an increase of nearly 50%.

What do these terms mean?

Digital Watermark: A subtle signal added to content so it is recognized as machine-generated without being noticed by the reader. Token: A small unit processed by a language model, which may be a word or part of a word, used to measure text length and generation speed. Open Source: Technology whose code is published publicly for inspection, use, and development.