谷歌 DeepMind 宣布，SynthID Detector 现已面向全球发布。该工具旨在帮助识别由 AI 生成的内容，包括使用 SynthID 合作伙伴（如 OpenAI、NVIDIA 和 Kakao）的 AI 工具创建的媒体内容，预计苹果公司也将很快提供支持。
在 SynthID Detector 推出之前，用户只能通过询问 Gemini 来识别 AI 生成的内容，但该功能仅限于谷歌 AI 生成的内容。如果要识别其他 AI 提供商的内容，则需要使用各自的检测工具。
SynthID.com 网站简化了这一流程，用户只要上传图片、音频或视频文件，便可以在所有 SynthID 合作伙伴（包括 OpenAI、NVIDIA、Kakao 以及即将加入的苹果）的系统中进行检测。该工具会给出“是”或“否”的结果。不过，谷歌也指出，当检测结果显示“无明确信号”时，阴性结果并不一定意味着该内容不是由 AI 生成的：
这并非一款通用的 AI 检测工具。我们仅能检测来自上述已采用 SynthID 技术的公司的媒体内容。我们正在不断地拓展合作伙伴关系，以实现对更多生成式媒体内容的检测。
具体来说，SynthID 无法检测使用 Meta 或微软 AI 生成的内容，因为它们采用了不同的水印技术。此外，大规模修改也会削弱 SynthID 的水印信号，从而导致误判。在这种情况下，用户需要采用更全面的方法，结合其他工具进行检测。尽管存在这些局限性，SynthID 仍然入选了《时代》杂志“2026 年最佳发明”榜单，该榜单认可了其相较于其他 AI 检测器在 AI 内容检测方面的创新方法。
针对这一公告，Reddit 用户 Chicano_Ducky 对该方法的有效性表达了负面看法：
[这]从一开始就没起过作用。文中提到，即使在他们自己模型生成的视频上，误报率也高得令人无法接受[……] 他们正试图将‘魔鬼’重新装回瓶子里，但根本做不到。
用户 Bulky-Employer 为这场讨论增添了新的视角，他强调：
创建水印移除节点非常简单。只需对图像进行一次低强度的降噪处理即可。你已经可以在 OpenAI 和 Gemini 生成的图像上测试这一点了。
对于那些想利用 AI 生成的图像进行欺诈的人来说，SynthID 只是个很小的障碍。它的存在对他们来说反而成了一件好事。
相较于对 AI 生成的内容进行水印标记，还有一种互补的方法是 C2PA 元数据。它记录文件的来源和编辑等历史信息，可以帮助验证其真实性。Harry G 在 LinkedIn 上指出，这是唯一“真正可扩展的方法”，因为“事后检测合成内容会演变成一场军备竞赛”。然而，C2PA 元数据在内容上传时可能会丢失，而且在截图中不会保留。
SynthID 由谷歌 DeepMind 于 2023 年推出，是一个针对 AI 生成的图像、音频、视频和文本进行水印加密以及检测的解决方案。它使用密钥改变音频频率概率分布或像素/token 模式，将不可见的水印直接嵌入到内容当中。
原文链接：https://www.infoq.com/news/2026/10/synthid-watermarking-detector/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Google DeepMind announces that Syndrid Detector is now facing the global publish. The tool is intended to help identify the content generated by AI, including media content created using the AI tool created by Syndrid partners (e.g. Open AI, NVIDIA and Kakao), and Apple is expected to provide support soon.
Before the SynthID Detector is launched, the user can only identify the content generated by AI by asking Gemini, but this feature is limited to the content generated by Google AI. If the content of other AI providers is to be identified, the respective testing tool is needed.
SynthID.com streamlines this process by allowing users to upload pictures, audio or video files to be tested in the systems of all SynthID partners, including Open AI, NVIDIA, Kakao and the incoming apples. The tool gives a "yes" or "no" result. However, Google also notes that when the test results indicate “no clear signal”, negative results do not necessarily mean that the content is not generated by AI:
This is not a generic AI testing tool. We can only detect media content from companies that have adopted SyndhID technology. We are constantly expanding partnerships to achieve the testing of more interactive media content.
Specifically, SyntID cannot detect content generated using Meta or Microsoft AI because they use different watermarking techniques. In addition, large-scale modifications weaken the watermarking signal of SyndhID, leading to miscalculation. In such cases, users need to adopt a more comprehensive approach, combining testing with other tools. Despite these limitations, SynthID was selected for the Time magazine's "Best Innovation 2026" list, which endorsed its innovative approach to AI content detection compared to other AI detectors.
In response to this announcement, Reddit user Chicano Ducky expressed a negative view of the validity of the method:
[This] didn't work from the beginning. It is mentioned that even on their own model videos, the rate of misreporting is unacceptably high [...] They are trying to reload the demon in the bottle, but they can't.
User Bulky-Employer added a new perspective to the discussion, emphasizing:
It is very simple to create watermarks to remove nodes. Just a low-intensity noise reduction treatment of the image. You can already test this on the images created by Open AI and Gemini.
For those who want to use the images generated by AI to commit fraud, SyntID is just a small obstacle. Its existence is good for them.
Another complementary method is C2PA metadata compared to watermarking of AI-generated content. It records historical information, such as the source and editing of the document, which can help verify its authenticity. Harry G. pointed out in LinkedIn that this is the only “really scalable method” because “the after-action detection of synthetic content will turn into an arms race”. However, C2PA metadata may be lost when their content is uploaded and will not be retained in the screenshot.
SynthID was launched by Google DeepMind in 2023 as a solution for watermarking and testing images, audio, video and text generated by AI. It changes the probability distribution of audio frequencies or the pixel/token model by using the key and embeds invisible watermarks directly into the content.
Original link: https://www.infoq.com/news/2026/10/synthid-watermarking-detector/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.