OpenAI to implement machine-readable watermarks for ChatGPT

OpenAI has announced plans to implement machine-readable watermarks for text generated by ChatGPT. This technical approach utilizes linguistic alternatives during the text generation process to create statistical patterns within the word stream.

The method involves manipulating the probability selection of tokens—the process where the model selects the next word. By subtly influencing this selection, a watermark is created that can be verified using a specific key. However, the effectiveness of detection and decryption remains probabilistic rather than absolute, depending on the length of the text and subsequent editing.

This move follows the implementation of the EU AI Act, which requires AI providers to include machine-readable labeling for AI-generated content to ensure transparency regarding the origin and production of digital material.