Google Deepmind Open Sources Its AI Text Watermarking Tool 

In an age where AI-generated content is increasingly common, concerns around authorship and authenticity are on the rise. The ease with which AI can create realistic text, images, and other forms of media has led to challenges distinguishing between human-created and AI-generated work. As a result, issues such as plagiarism, copyright infringement, and the spread of misinformation have gained prominence, raising the need for tools that can verify content origin.

Google DeepMind, in collaboration with Hugging Face, developed SynthID Text, a watermarking technology created to make AI-generated text easier to identify. The tool is now open source and is accessible through the Google Responsible GenAI Toolkit. 

Last year, Google DeepMind launched a watermarking solution for images, and it has since expanded this technology to cover AI-generated videos as well. In May this year, the company unveiled SynthID and announced it would integrate SynthID into its Gemini app and online chatbots and also made the tool freely accessible on Hugging Face, an open-source repository for AI models and datasets.

Currently, SynthID for text works exclusively with content generated by Google’s models. However, open-sourcing it is intended to broaden its compatibility with various tools and platforms.

“Now, other [generative] AI developers will be able to use this technology to help them detect whether text outputs have come from their own [large language models], making it easier for more developers to build AI responsibly,” says Pushmeet Kohli, the vice president of research at Google DeepMind. 

Specialized LLMs can play a vital role in GenAI rollout

SynthID Text, detailed in a paper published in Nature on Oct. 23, works by encoding a watermark into AI-generated text in a way that helps identify if a specific language model produced it. 

Importantly, this process leaves the LLM’s core functionality untouched and maintains the quality of the generated content, ensuring that the watermarking doesn’t alter the user experience or reduce text fluency.

It’s important to clarify that SynthID is not designed to detect text generated by any language model. Instead, it specifically watermarks the output from a designated LLM. 

Large language models break down language into “tokens,” which can be characters, words, or phrases. They predict the next token by assigning percentage scores based on their likelihood of following the previous tokens; higher scores indicate a greater chance of being selected.

SynthID works by watermarking AI-generated content by adjusting the probability of specific tokens to embed a watermark. It detects this watermark by analyzing the probability scores of words in both watermarked and unwatermarked texts.

The approach taken by Deepmind researchers is not new. OpenAI, the company behind ChatGPT, is testing a similar method. Additionally, significant contributions have been made by Scott Aaronson, a computer scientist at the University of Texas at Austin, as well as John Kirchenbauer and his colleagues at the University of Maryland, who have developed both watermarking and detection algorithms. 

DeepMind's team has enhanced this concept by successfully implementing it at scale while keeping computational costs low. According to the SynthID Text developers, the tool is “the first deployment of a generative text watermark at scale.” 

In a study involving approximately 20 million chatbot responses, DeepMind researchers compared those marked with SynthID to unwatermarked ones. The company states that the analysis showed users did not perceive any differences in the quality or usefulness of the text, indicating that SynthID can maintain output integrity while providing essential identification features. 

While SynthID Text offers key benefits, it also has notable limitations. The watermark can resist minor tampering, but it may fail against significant alterations, particularly if the content is heavily rewritten or translated. Additionally, its reliability diminishes when AI-generated text is translated into different languages.

The tool’s effectiveness will also be challenged with prompts requiring factual information, as this restricts SynthID’s ability to predict the next word without changing the meaning. Other potential limitations include a reduced scope of detection for highly modified text and a lack of context sensitivity.

The need for watermarking AI-generated text is clear. However, they must gain users' trust for such tools to be effective. Such watermark tools must be watertight to be reliable. DeepMind's decision to make the SynthID-Text model publicly available marks a significant step forward, but the technology is still in its early stages. Further research and development would be needed to ensure the robustness and reliability of watermarking methods.