Google DeepMind has unveiled the SynthID Text open-source code, a tool to create watermarks indicating that a message is AI-generated for future verification purposes. The concept of inserting watermarks may seem straightforward visually in images, as humans can easily distinguish them, but inserting watermarks in text messages poses a different challenge altogether.
SynthID Text operates on the principle that the Language Model (LLM) acts as a “word dispenser,” predicting the next word based on probability. The token with the highest probability is selected as the output for users to see. The tool alters the “probability” values just before the LLM dispenses words, ensuring different results without the need for retraining the model. While the outcomes may vary slightly, SynthID strives to maintain the quality of results close to the original LLM output.
Results generated through SynthID may contain certain words with adjusted probability weights by the tool, which can be directly manipulated by the SynthID software, allowing LLM model creators to understand how weights are adjusted and configure the AI content detection pattern accurately.
DeepMind mentions that the SynthID Text technique is effective for short text messages consisting of at least three sentences, with higher accuracy in longer text messages. However, if the text is later modified (e.g., by humans), the detection accuracy may decrease.
The SynthID technique has already been applied to Gemini-generated texts, with Google DeepMind’s research team testing nearly 20 million marked and unmarked texts to assess result quality through user feedback. The difference in quality was found to be negligible (0.01%-0.02% difference in scores).
The SynthID Text library is now open-source on GitHub under the Apache 2.0 license, with Hugging Face incorporating it into their Transformers model platform.
TLDR: Google DeepMind introduces SynthID Text, a tool for watermarking AI-generated text for future verification, ensuring content quality and integrity.
Leave a Comment