Google is making SynthID Textual content, its technology that allows developers to watermark and detect written text using generative AI models, generally available.
SynthID Textual content can be downloaded from the Face Hugging AI Platform and Google's updated Responsible GenAI Toolkit.
“We are open source our SynthID Textual content watermarking tool,” the company wrote in an email on X. “Freely available to developers and businesses, it will help them identify their AI-generated content.”
So how does it work?
Given a message like “What is your favorite fruit?”, text generation models predict which “token” is likely to follow another: one token at a time. Tokens, which can be a single character or a word, are the building blocks that a generative model uses to process information. A model assigns each possible token a score, which is the percentage probability that it will be included in the output text. SynthID Textual content inserts additional information into this token distribution by “modulating the probability of tokens being generated,” Google says.
“The final pattern of scores for the model's word choices combined with the adjusted probability scores is considered the watermark,” the company wrote in a blog post. “This pattern of scores is compared to the expected pattern of scores for watermarked and unwatermarked text, helping SynthID detect whether the text was generated by an AI tool or whether it could come from other sources.”
Google claims that SynthID Textual content, which has been integrated with its Gemini models since this spring, does not compromise the quality, accuracy, or speed of text generation, and works even with trimmed, paraphrased, or modified text.
But the company also admits that its watermarking approach has limitations.
For example, SynthID Textual content does not work as well with short texts or texts rewritten or translated from another language, or with answers to objective questions. “In responses to factual prompts, there is less opportunity to adjust token distribution without affecting factual accuracy,” the company explains. “This includes messages like 'What is the capital of France?' or queries where little or no variation is expected, such as 'recite a poem by William Wordsworth'.”
Google is not the only company working on AI text watermarking technology. OpenAI has been researching watermarking methods for years, but delayed its release due to technical and commercial considerations.
Watermarking techniques, if widely adopted, could help turn the tide on inaccurate but increasingly popular “AI detectors” that false flag essays written in a more generic voice. But the question is, willpower will be widely adopted, and will one standard or technology prevail over others?
There may soon be legal mechanisms that bind developers. China's government has introduced mandatory watermarks on AI-generated content and the state of California is looking to do the same.
The situation is urgent. According to a report from the European Union Law Enforcement Agency, 90% of online content could be synthetically generated by 2026, creating new challenges for law enforcement around disinformation, propaganda, fraud and deception.