谷歌照片正在为其Magic Editor的生成式AI功能编辑过的照片添加数字SynthID水印。此功能旨在帮助人们快速识别使用Magic Editor中的“reimagine”工具修改过的图像。SynthID是由谷歌DeepMind团队创建的水印系统，它将数字元数据标签直接嵌入到图像、视频、音频和文本中，以识别它们是否是使用AI工具创建或修改的。然而，SynthID并不明显改变图像，需要专门的AI检测工具才能识别，且某些编辑可能太小而无法被检测到。专家认为，仅靠水印不足以可靠地验证大规模的AI生成内容，需要多种方法协同工作。

🖼️谷歌照片为Magic Editor编辑过的照片添加SynthID水印，旨在方便用户识别AI修改过的图片。

🤖SynthID是由谷歌DeepMind团队开发的水印系统，通过嵌入数字元数据来识别AI生成或修改的内容，目前已应用于Google的Imagen文本到图像模型生成的图片。

🔍SynthID水印不会明显改变图像，需要通过谷歌的“About this image”功能中的AI检测工具进行识别，但部分修改幅度较小的图片可能无法被有效检测。

⚠️专家指出，仅靠水印技术不足以完全验证AI生成内容的真实性，需要结合多种方法才能更有效地识别和验证。

Google Photos is adding its digital SynthID watermarks to photos that have been edited using the Magic Editor’s generative AI feature. The new feature is rolling out “this week” according to Google, and is intended to make it easier for people to quickly identify images that have been manipulated using the “reimagine” tool in Magic Editor.

SynthID is a watermarking system created by Google’s DeepMind team that embeds a digital metadata tag directly into images, video, audio, and text to identify if they were created or altered using AI tools. The watermark is already being applied to images that are entirely AI-generated using Google’s Imagen text-to-image model. Similar watermarking systems have also been developed by other companies, like the Content Credentials that Adobe applies to works created or edited using its suite of Creative Cloud apps.

My colleagues have previously covered how easy it is to convincingly edit photos using the reimagine feature available on Google’s Pixel 9 phones by just describing what you want it to add. While AI editing tools aren’t inherently nefarious, Magic Editor was happy to generate things like crashed helicopters, drug paraphernalia, and even corpses in photos with no obvious way to tell the manipulations were AI-generated. (Google did start tagging AI-edited images in the file descriptions on Google Photos last October, at least.)

That’s a problem the new watermarking update is attempting to solve, but there are a few issues. For one, SynthID doesn’t visibly alter the image it’s attached to and can be identified using a dedicated AI detection tool that’s a part of Google’s “About this image” feature. Google also says that some edits made using the Magic Editor reimagine “may be too small for SynthID to label and detect.” 

While watermarking tech can help if it does catch manipulations, some experts in this field agree that watermarking alone won’t be effective enough to reliably authenticate AI-generated content at scale, and will require a varied suite of different approaches working in tandem.