Google Releases an AI Tool That Watermarks AI Generated Text (25.10.24)

Authored by: Ms Tanima Bhatia

Key Highlights:

Open-Source Availability & Integration: SynthID Text, Google’s AI-generated text watermarking tool, is now open-source on Hugging Face and integrated into Gemini models since Spring 2024.
Watermarking Mechanism: It subtly alters token probabilities to embed a watermark without affecting text quality, with detection based on identifying these token patterns.
Performance & Limitations: Proven resistant to light edits but less effective against heavy paraphrasing, translation, or factual content generation. Successfully tested across 20 million chatbot responses with no impact on creativity or speed.

Google’s SynthID Text is a groundbreaking tool that enables developers to watermark and detect AI-generated text, with the goal of promoting responsible use of generative AI. The technology has been integrated into Google’s Gemini models and is now open-sourced through Hugging Face and Google’s Responsible GenAI Toolkit.

How SynthID Text Works

SynthID Text modifies the text generation process by embedding an invisible watermark within the token probabilities. Tokens, essentially characters, words, or parts of phrases, are the building blocks used by large language models (LLMs). Each token is assigned a probability score indicating how likely it is to be generated next.

The watermark works by modulating the likelihood of certain tokens being selected. This slight but intentional adjustment creates a detectable pattern within the token distribution. When reviewing generated text, SynthID compares the token patterns against expected watermarked distributions to determine whether the text was AI-generated.

Performance and Limitations

Google DeepMind claims that the watermark does not affect the text’s quality, creativity, or speed. In fact, after deploying SynthID across 20 million Gemini chatbot interactions, user ratings confirmed that watermarked and non-watermarked responses were indistinguishable in quality.

However, SynthID has limitations. It remains somewhat effective against paraphrasing or cropping, but the watermark weakens if the text undergoes translation or extensive rewriting. Additionally, the method struggles with highly factual content, like trivia or coding, where fewer opportunities exist to subtly alter token probabilities.

Impact and Future Adoption

Google’s decision to open-source SynthID aims to encourage widespread adoption among developers and AI companies. Tools like this are crucial in identifying AI-generated text, helping address issues such as misinformation and AI-driven plagiarism. Pushmeet Kohli, a Google DeepMind researcher, hopes that open sourcing will foster community-driven improvements and encourage other companies to adopt similar watermarking practices.

Still, experts caution that watermarking isn’t foolproof. Researchers have demonstrated that determined attackers can remove or “scrub” watermarks, posing challenges to their long-term effectiveness. Additionally, coordinated industry efforts are needed to ensure that watermarks become standardized across platforms.

Conclusion

SynthID Text represents a major step toward responsible AI, providing a strong mechanism for watermarking AI-generated content. While it isn’t a panacea for all risks, it contributes to the growing ecosystem of safeguards. The open-source release allows developers to explore and refine the technology further, ensuring that watermarking becomes a foundational component in the responsible deployment of generative AI models.