Python experiments to add, and detect watermarks in natural language text. I originally outlined this idea in a suggestion to OpenAI in September 2022.
With my initial algorithm, some of the original inspiration came from markov chain based text steganography by H Moraldo, 2014 using n-grams (before LLMs offered reliable probabilities for next token)
- Kirchenbauer et al, 2023
- A more robust implementation (including Reed Solomon error correction) is exciting, available here