The move that will make you think twice before using artificial intelligence

OpenAI begins embedding invisible watermarks in ChatGPT texts following the European AI Act. The technology, textGrain, subtly influences word selection to create a unique statistical signature. The move will initially apply only to users in the European Union, but signals a global regulatory trend.

OpenAI is beginning to embed invisible digital watermarks in texts produced by ChatGPT and the Codex development environment, following Article 50 of the European Union's Artificial Intelligence Act (EU AI Act). The technology, called textGrain, subtly influences the distribution of words the model chooses, using a secret key, and generates a unique statistical signature that enables computerized identification of synthetic content. Performance tests on the GPT-6 Astra model showed that the watermarking does not compromise answer quality in programming, automation, and professional knowledge tests. However, detection effectiveness depends on paragraph length and domain: in psychology detection with 400-token paragraphs it stood at 95%, but dropped to 80% in shorter paragraphs. Simple human editing, such as replacing 10% of words with synonyms, reduced detection from 92% to 66%. At this stage, the company will not release the detector to the general public due to concerns about false positives. The move will initially apply only to users in the European Union, but signals a global trend of regulation and transparency.