Google has announced the general availability of SynthID Text, a tool that allows developers to watermark and detect text generated by AI models. This technology can be accessed via the AI platform Hugging Face and Google’s updated Responsible GenAI Toolkit.
In a post on X, Google shared, “We’re open-sourcing our SynthID Text watermarking tool. It will be freely available to developers and businesses, enabling them to identify their AI-generated content.”
So, how does SynthID Text work?
When given a prompt like “What’s your favorite fruit?,” text-generating models predict the next likely “token,” which can be a character or word, one at a time. These tokens are the fundamental elements that generative models use to process information. Each potential token is assigned a score representing the likelihood of its inclusion in the output. SynthID Text enhances this token distribution by adjusting the probability of certain tokens being generated.
As Google explains, “The final pattern of scores, combining the model’s word choices with the adjusted probabilities, constitutes the watermark.” This score pattern is then compared to expected patterns for both watermarked and unwatermarked text, enabling SynthID to determine whether the text was generated by an AI tool or derived from other sources.
Google asserts that SynthID Text, integrated with its Gemini models since spring, does not compromise text generation quality, accuracy, or speed, and it remains effective even with cropped, paraphrased, or modified text. However, the company acknowledges some limitations of its watermarking approach.
Read Also: Native Union’s Smart Little Combo Charger Has Just Secured A Place In My Travel Bag.
For instance, SynthID Text struggles with short text, rewritten content, translated text, and factual responses. “For factual prompts, there are fewer chances to adjust the token distribution without affecting accuracy,” the company noted. This includes questions like “What is the capital of France?” or requests for specific poems, where minimal variation is expected.
Google is not alone in developing AI text watermarking technologies; OpenAI has also explored watermarking methods but has delayed their release due to technical and commercial factors.
Widespread adoption of text watermarking could help combat inaccurate “AI detectors” that mistakenly flag more generic writing as AI-generated. However, it remains uncertain whether a single standard or technology will dominate.
Legal measures may soon prompt developers to take action. China has introduced mandatory watermarking for AI-generated content, and California is considering similar regulations.
The urgency of this issue is highlighted by a report from the European Union Law Enforcement Agency, which predicts that by 2026, 90% of online content could be synthetically generated, presenting new challenges for law enforcement regarding disinformation, propaganda, and fraud. An AWS study indicates that nearly 60% of all web sentences may already be AI-generated, largely due to the widespread use of AI translators.
