The Long-Awaited Answer to “How to Identify AI-Generated Text” Is Here: Google’s SynthID Text Watermark Tool 

Generative AI has proven to be amazing at creating text – so good that it has been impossible to tell for sure whether the text has been written by a human or generated by AI. As we’ve seen, this has led to all sorts of problems, from widespread misinformation to students being falsely accused of plagiarism. Now, researchers at Google’s Deep Mind labs think they have come up with a solution: SynthID – a tool that writes and reads text watermarks – to finally have an answer for how to detect AI-generated text.

A paper published October 23, 2024, titled “Scalable watermarking for identifying large language model outputs” discusses a new watermarking scheme called SynthID.According to these researchers, SynthID Text successfully distinguishes AI-generated text from human-written text and does it at scale.

SynthID was tested across multiple large language models (LLMs) and showed improved detectability compared to existing methods. It was also evaluated in a live experiment with nearly 20 million responses using Google’s Gemini and Gemini Advanced chatbots, confirming that it does not degrade text quality.

In the simplest terms, when the AI writes something, SynthID adds a hidden “signature” to the text. It slightly adjusts its word choices to include this signature. It’s like choosing specific words that fit a pattern only the tool can recognize, but it still makes sense to us humans since text still reads naturally and smoothly. Later, if someone wants to check if the text was written by AI, they can use a special detector that looks for the hidden signature in the text.

The authors of the paper note that “generative watermarks require coordination between actors running the LLM text generation services to apply the watermark”. They also have concern over open-source models, since enforcing watermarks on decentralized deployments will be difficult.

Finally, bad actors have been trying to find ways around generative text watermarking; further research will be needed to determine how SynthID compares with other watermarking schemes in terms of vulnerability.

Despite the noted limitations, the development of watermarking schemes like SynthID will go a long way toward improving AI text generation through:

The code to generate and detect text watermarks with SynthID is available right here on GitHub.

Understanding how to detect AI-generated text is crucial. Text watermark tools like Google’s SynthID are at the forefront of this effort, providing innovative solutions so we can better navigate the complexities of the digital landscape and maintain trust in the content we encounter. SynthID can be a big step in the right direction toward a safer, more compliant and more transparent AI future.