Researchers introduce a cryptographically-inspired method for embedding undetectable watermarks in large language models like GPT-4.
These watermarks can only be detected with a secret key, making it computationally intractable to distinguish watermarked outputs from original model outputs.