Google has unveiled a new tool called SynthID Text, designed to mark and identify texts generated by AI models. This technology will not compromise the quality or speed of content creation and will be freely available for developers and companies.
How SynthID Text Works
SynthID Text operates by modulating the probabilities used to generate text. AI models predict the next “token” (a word or symbol) based on probability distributions. By subtly adjusting these probabilities, the tool embeds a watermark, enabling the detection of AI-generated content. As Google explains, the “pattern of selected probabilities, combined with adjusted estimates, will serve as the watermark.”
Although SynthID Text helps detect AI-created text, slight changes such as rephrasing or trimming might reduce its effectiveness. The model also has some limitations with highly specific or unambiguous queries, as it leaves less room to adjust tokens without affecting factual accuracy.
Industry Adoption and Regulation
Google first integrated SynthID Text into its Gemini model in spring, ensuring the generation speed and accuracy remain unaffected. However, Google isn’t the only player in this space—OpenAI was also working on watermarking technology but postponed its launch due to technical and business challenges.
If broadly adopted, technologies like SynthID Text could mitigate the risks posed by inaccurate AI detection tools, which sometimes misidentify student essays or other original content as AI-generated, notes NIX Solutions. Governments are already taking action: China requires AI content to be labeled, and California is planning to introduce similar regulations.
As watermarking technology evolves, we’ll keep you updated on its impact and potential global adoption.