Feature Request: Publicly Verifiable Watermark Detection with SynthId

[rifkiamil]

I am writing to request a feature that would significantly enhance the trustworthiness and accountability of content generated with synthid. Specifically, I propose the development and release of a mechanism for publicly verifiable watermark detection without requiring access to private keys or sensitive model parameters.

Problem: The increasing sophistication of generative AI raises concerns about misinformation, deepfakes, and the ability to attribute content to its source. While watermarking is a promising approach to address these issues, current methods often rely on private keys held by the model creators. This limits the ability of independent parties (e.g., the public, journalists, researchers, government agencies, platform owners) to verify the origin of content.

Proposed Solution: I believe it's crucial to enable independent verification of watermarks. This could potentially be achieved through techniques like zero-knowledge proofs (ZKPs) or other cryptographic methods that allow someone to prove that a watermark is present without revealing the watermark itself or the key used to generate it.

Use Cases:

• Public Transparency: Allow anyone to check if a piece of text, image, or video was generated by a model using synthid, promoting transparency and combating misinformation.
• Journalistic Integrity: Enable journalists to verify the authenticity of sources and identify potentially manipulated content.
• Government Oversight: Provide government agencies with tools to detect the misuse of AI for malicious purposes.
• Platform Moderation: Help platform owners (e.g., social media companies) identify and flag content generated by known, safe LLMs versus potentially harmful or untrusted sources.
• Research and Auditing: Allow researchers to audit LLM behavior