Google Unveils SynthID Text

Google DeepMind and Hugging Face have released SynthID Text, which embeds and detects watermarks in text to enhance the trustworthiness of AI-generated content, integrating it into Transformers v4.46.0.

This technology embeds a watermark during the generation process that is imperceptible to humans but identifiable by a trained model, without affecting the generation quality or operation of the LLM.

Key Technology and Features:

• How It Works: Uses a pseudorandom function (g-function) to encode the watermark during the generation process.
• Versatility: Provided as a logits processor compatible with any LLM via the model.generate() API, without requiring modification of the model.
• Detection: Watermarked text can be identified through a dedicated classifier.

Configuration Parameters:

• keys: A list of random integers used for watermark security, which must be managed securely to prevent others from replicating the watermark.
• ngram_len: Controls the balance between the watermark's detection performance and robustness (default value 5).