Google's AI watermark system

Introduction

Google has announced the launch of SynthID, a watermarking technology that puts a stamp on all AI-created content regardless of its type – text, picture, video or even music. Small corrections or big alterations, invasive edits such as cropping, altering the video speed, compressing, and color correction do not affect the core message of the video which is the watermark.

How SynthID works

Developer DeepMind considers SynthID as a new watermarking technology denoting an AI content creator, text, or a still picture or video embedded with rather complex unique markers that are hard to erase. However complicated the editing may be — compression, cropping, or color adjustments, for example — the watermark endures.

SynthID forms words based on probabilities to replace text, combine visuals into a solid however pixelated image or video, and transfer sequences without loss of sound quality. The growing prevalence of deep fake and misinformation encourage the need for SynthID, which will work to enable the accurate identification of AI material from original content.

This technology is currently in beta on Google’s AI products and is also open-sourced via the Google Responsible Generative AI Toolkit. With this, developers can experiment and explore further with other collaborators such as Hugging Face. SynthID holds a promise in helping people discern genuine information from the content generated by AI, supporting further transparency in its use. For more information, see DeepMind’s official page about SynthID.

The SynthID powerful

What makes SynthID powerful is that it watermarks at the word selection level, whereby only three sentences are needed to catch it. This is practical for those who borrow AI to write essays but use AI to humanize the text so that it looks more real.

Text: It embeds a watermark during word selection, with just three sentences needed to catch it. This is practical for those who borrow AI to write essays but still use AI to humanize the text for a more authentic look.

Images/Videos: The watermark is embedded in every pixel and frame. Whether the content has been cropped, compressed, or changed altogether, that watermark goes entirely undisturbed.

Audio: The watermark goes in the spectrogram of the sound waves. Not even MP3 compression or speed change appears to affect it!

At a time when deepfakes, fake news, and other scams keep emerging left and right, SynthID rapidly sifts the real content from AI-generated ones in order to protect against AI misuse effectively.

SynthID is currently in beta on Google’s AI products and has open-sourced the tool through the Google Responsible Generative AI Toolkit. Besides that, it partners with Hugging Face, where developers find an open playground to further this concept.

This article was created by AI.