SynthID Detector Now Open to the Public | Anyone Can Check AI-Generated Images
機械翻訳 / Machine-translated
A
AI Friends
4日前・15分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever looked at a photo on social media and thought, "Did AI make this?" Google has made a tool available to everyone that can help answer that question. This article breaks down how to use it and what its limitations are.
On October 7, 2026, Google DeepMind announced on its official blog that "SynthID Detector" is now available to everyone.
SynthID Detector is a website designed to identify images, videos, and audio created by AI. By uploading a file, you can check whether it was made by Google or one of its partner companies' AI systems.
In fact, this tool was announced some time ago. When GIGAZINE reported on it in May 2025, it was a limited version available only to journalists and researchers on a waitlist.
With this public release, those restrictions have been lifted. The English version of the SynthID Detector site is now accessible from anywhere in the world.
SynthID is a technology Google launched in 2023. At the moment AI creates content, it embeds an invisible digital watermark (a hidden stamp) that cannot be perceived by the human eye or ear.
According to GIGAZINE's 2025 article, for images the watermark is embedded directly into the pixels themselves. Because of this, the watermark is said to persist even if metadata (attached data such as shooting information) is removed or the image is edited.
Google states that it has embedded watermarks in more than 18 billion images and videos, and in audio totaling 240,000 years of playback time.
In other words, much of the AI-generated content produced through Google's services already carries a watermark. This new tool serves as the gateway for reading those watermarks.
According to GIGAZINE's 2026 article, the process works as follows:
You can log in using a ChatGPT account. The fact that an OpenAI account works with a Google tool may reflect the deepening partnership between the two companies.
Results are displayed in a format such as "SynthID detected." If a partner company's AI is identified, the source is also shown — for example, "This media was created or edited with OpenAI."
Pricing is not mentioned in the GIGAZINE article. It was also not stated in Google's official blog.
The tool can identify content created by Google's own AI and by partner companies' AI. The official blog names OpenAI, NVIDIA, and Kakao as partners. Apple is listed as "coming soon."
Here are the results a GIGAZINE reporter obtained when testing several images:
Pay close attention to the third result. After editing, the watermark may remain but the source may no longer be identifiable.
SynthID is not the only way to identify AI-generated content. C2PA is another prominent method. According to a Sesame Disk article, the two serve different roles.
C2PA records the history of who created content and how, as signed metadata. However, its weakness is that this data is easily lost through screenshots, re-encoding, or uploading to social media.
SynthID embeds a signal at the time of generation. While the signal tends to survive even after metadata is lost, the same article notes that the signal can be weakened to undetectable levels through specialized processing.
For text, the article also states that rephrasing with a different AI removes the statistical signal, significantly reducing detection accuracy.
The article recommends a "layered defense" using both C2PA and SynthID for images and video — the idea being that neither alone is sufficient.
This is a useful tool, but it is not all-powerful. GIGAZINE's article identifies three key limitations:
For example, imagine a PR representative at a company checking product photos received from a business partner. Even if the result says "Not detected," that is no reason to feel reassured — the image may have been created with an AI that does not use SynthID watermarks.
Conversely, a "Detected" result is a strong clue that AI was involved. However, whether the photo is a malicious fake requires separate verification.
It is best to treat results as a clue rather than a verdict. Combine this tool with other methods, such as following up with the original poster or tracing the source through a reverse image search.
Many people are probably wondering whether they can use this from Japan. According to GIGAZINE, the English version is available worldwide — meaning it should be accessible from Japan, though the interface is in English only.
Whether a Japanese-language version will be offered was not mentioned in any article we were able to confirm.
In Japan, concerns about deepfakes (convincingly realistic fake videos and audio) are growing. For journalists who want to verify whether a news image is authentic, or teachers who want to check images distributed at school, having a free entry point to test this matters a great deal.
At the same time, many images created by domestic AI or open-source overseas AI do not carry watermarks. This tool alone cannot verify every image circulating in Japan.
Pricing was not mentioned in any article we were able to confirm. Please check the site's own guidance before using it.
Only an English version is available. Whether a Japanese version will be released has not been confirmed.
No. It may have been created by an AI that does not use SynthID.
The official blog states that images, videos, and audio are supported. Text is not included in the official blog's description for this release.
OpenAI is listed as a partner company. In GIGAZINE's testing, an image created with ChatGPT was displayed as "created or edited with OpenAI."
Start by uploading an AI-generated image you have on hand and see what results you get.
This article is a cross-post from AI Friends.
