SynthID injects era-specific nuances by dynamically adjusting the probability of token generation, as explained by Kohli.
SynthID detects watermarks and determines whether AI-generated text was used by comparing the predicted probability scores for phrases between watermarked and unwatermarked content.
Google’s DeepMind division found that embedding the SynthID watermark had no discernible impact on the quality, precision, originality, or tempo of generated written output. The study’s findings were based on a comprehensive trial conducted with SynthID’s Gemini products, which had been implemented for thousands of users. Customers are empowered to evaluate the quality of AI model responses by providing a simple thumbs-up or thumbs-down ranking.
Kohli and his team analyzed the scores of approximately 20 million chatbot responses, both watermarked and unwatermarked, to gain insights. The research found that customers failed to perceive any significant difference in terms of quality and utility between the two options. The outcomes of this experiment are documented in a published paper currently available. Currently, SynthID’s functionality is limited to processing textual content generated by Google’s models; however, the plan is that open-sourcing it will expand its compatibility with a broader range of tools?
SynthID does produce other limitations. Despite its immunity to minor tampering, the watermark’s reliability wavered significantly when confronted with AI-generated text that had been extensively rewritten, translated, or manipulated in some way. While it’s typically unreliable when responding to requests for verifiable information like the capital of France? Because the scarcity of options constrains the likelihood of incorporating a coherent phrase within a sentence without compromising its essence.
